Two optimizations for the resolve_qualified_name hot path:

1. from_two_parts: directly construct inline SegmentsStack for the common
   case (≤8 segments), bypassing the SegmentsVec API overhead (match on
   variant, capacity checks, spill handling on each extend_from_slice).

2. resolve_qualified_name: extract attribute tail segments directly from
   the AST for single and double attribute access (e.g., os.path,
   os.path.join), avoiding the cost of UnqualifiedName::from_expr
   entirely for these common cases.

Walltime benchmarks show -4.3% to -5.5% improvement on large files.


Co-Authored-By: CodSpeed <no-reply@codspeed.io>

Two optimizations targeting physical line processing:

1. bidirectional_unicode: Skip the contains() check for ASCII-only lines,
   since all bidirectional Unicode characters are non-ASCII. The is_ascii()
   check is SIMD-accelerated and handles the common case immediately.

2. LineWidthBuilder::add_str: For ASCII text, iterate bytes directly
   instead of chars, avoiding per-character unicode_width lookups. ASCII
   characters always have width 1 (except tabs and newlines which are
   handled explicitly).

Walltime benchmarks show -1% to -3% additional improvement across all
test files, with the largest gains on large/dataset.py.


Co-Authored-By: CodSpeed <no-reply@codspeed.io>

- Add super-fast path in `add_str` using `memchr3` to skip byte-by-byte
  iteration for ASCII text without tabs/newlines/CR (just add byte length)
- Add tighter early exit in `Overlong::try_from_line`: for any text,
  display width ≤ UTF-8 byte length (except tabs), so lines with byte
  length ≤ limit and no tabs cannot be overlong
- Hoist `indented_form_feed` rule check outside the per-line loop in
  `check_physical_lines`


Co-Authored-By: CodSpeed <no-reply@codspeed.io>