Add API documentation

• PR#423: add Doxygen-based API documentation, now hosted in https://rapidyaml.readthedocs.io/!
• It uses the base doxygen docs, as I couldn't get doxyrest or breathe or exhale to produce anything meaningful using the doxygen groups already defined in the source code.

Error handling

Fix major error handling problem reported in #389 (PR#411):

• The NodeRef and ConstNodeRef classes are now conditional noexcept using RYML_NOEXCEPT, which evaluates either to nothing when assertions are enabled, and to noexcept otherwise. The problem was that these classes had many methods explicitly marked noexcept, but were doing assertions which could throw exceptions, causing an abort instead of a throw whenever the assertion called an exception-throwing error callback.
• This problem was compounded by assertions being enabled in every build type -- despite the intention to have them only in debug builds. There was a problem in the preprocessor code to enable assertions which led to assertions being enabled in release builds even when RYML_USE_ASSERT was defined to 0. Thanks to @jdrouhard for reporting this.
• Although the code is and was extensively tested, the testing was addressing mostly the happy path. Tests were added to ensure that the error behavior is as intended.
• Together with this changeset, a major revision was carried out of the asserting/checking status of each function in the node classes. In most cases, assertions were added to functions that were missing them. So beware - some user code that was invalid will now assert or error out. Also, assertions and checks are now directed as much as possible to the callbacks of the closest scope: ie, if a tree has custom callbacks, errors within the tree class should go through those callbacks.
• Also, the intended assertion behavior is now in place: no assertions in release builds. Beware as well - user code which was relying on this will now silently succeed and return garbage in release builds. See the next points, which may help.
• Added new methods to the NodeRef/ConstNodeRef classes:

  /** Distinguish between a valid seed vs a valid non-seed ref. */
  bool readable() const { return valid() && !is_seed(); }
  
  /** Get a child by name, with error checking; complexity is
   * O(num_children).
   *
   * Behaves as operator[](csubstr) const, but always raises an
   * error (even when RYML_USE_ASSERT is set to false) when the
   * returned node does not exist, or when this node is not
   * readable, or when it is not a map. This behaviour is similar to
   * std::vector::at(), but the error consists in calling the error
   * callback instead of directly raising an exception. */
  ConstNodeRef at(csubstr key) const;
  /** Likewise, but return a seed node when the key is not found */
  NodeRef at(csubstr key);
  
  /** Get a child by position, with error checking; complexity is
   * O(pos).
   *
   * Behaves as operator[](size_t) const, but always raises an error
   * (even when RYML_USE_ASSERT is set to false) when the returned
   * node does not exist, or when this node is not readable, or when
   * it is not a container. This behaviour is similar to
   * std::vector::at(), but the error consists in calling the error
   * callback instead of directly raising an exception. */
  ConstNodeRef at(size_t pos) const;
  /** Likewise, but return a seed node when pos is not found */
  NodeRef at(csubstr key);

• The state for NodeRef was refined, and now there are three mutually exclusive states (and class predicates) for an object of this class:
  • .invalid() when the object was not initialized to any node
  • .readable() when the object points at an existing tree+node
  • .is_seed() when the object points at an hypotethic tree+node
  • The previous state .valid() was deprecated: its semantics were confusing as it actually could be any of .readable() or .is_seed()
• Deprecated also the following methods for NodeRef/ConstNodeRef:

  RYML_DEPRECATED() bool operator== (std::nullptr_t) const;
  RYML_DEPRECATED() bool operator!= (std::nullptr_t) const;
  RYML_DEPRECATED() bool operator== (csubstr val) const;
  RYML_DEPRECATED() bool operator!= (csubstr val) const;

• Added macros and respective cmake options to control error handling:
  • RYML_USE_ASSERT - enable assertions regardless of build type. This is disabled by default. This macro was already defined; the current PR adds the cmake option.
  • RYML_DEFAULT_CALLBACK_USES_EXCEPTIONS - make the default error handler provided by ryml throw exceptions instead of calling std::abort(). This is disabled by default.
• Also, RYML_DEBUG_BREAK() is now enabled only if RYML_DBG is defined, as reported in #362.
• As part of PR#423, to improve linters and codegen:
  • annotate the error handlers with [[noreturn]]/C4_NORETURN
  • annotate some error sites with C4_UNREACHABLE_AFTER_ERR()

More fixes

• Tree::arena() const was returning a substr; this was an error. This function was changed to:

  csubstr Tree::arena() const;
  substr Tree::arena();
  

• Fix #390 - csubstr::first_real_span() failed on scientific numbers with one digit in the exponent (PR#415).
• Fix #361 - parse error on map scalars containing : and starting on the next line:

  ---
  # failed to parse:
  description:
    foo:bar
  ---
  # but this was ok:
  description: foo:bar

• PR#368 - fix pedantic compiler warnings.
• Fix #373 - false parse error with empty quoted keys in block-style map (PR#374).
• Fix #356 - fix overzealous check in emit_as(). An id may be larger than the tree's size, eg when nodes were removed. (PR#357).
• Fix #417 - add quickstart example explaining how to avoid precision loss while serializing floats (PR#420).
• Fix #380 - Debug visualizer .natvis file for Visual Studio was missing ConstNodeRef (PR#383).
• FR #403 - install is now optional when using cmake. The relevant option is RYML_INSTALL.

Python

• Fix #428/#412 - Parse errors now throw RuntimeError instead of aborting.

Thanks

• @Neko-Box-Coder
• @jdrouhard
• @dmachaj