fix(jit): retry jitdump read when file grows (#553)

fix(jit): retry jitdump read when file grows

RuntimeSymbolLookup::get_or_insert_jitdump freezes a (pid, path) pair
for the rest of the export cycle as soon as one sampled PC misses the
symbol map after a successful jitdump read. The intent was to avoid
reparsing the file for every subsequent miss in the same cycle.

The side effect is that a single unlucky first sample -- fired while
the runtime is still flushing jit-<pid>.dump (Julia, .NET) -- produces
zero symbolised JIT frames for the entire run when the scenario only
exports once. This matches the flaky failure seen on the
prof-correctness Julia scenario.

Add a soft-miss retry path: every 64 soft misses on a frozen jitdump,
stat the file via the existing get_file_inode helper. If size or inode
changed since the last read, unfreeze and reparse; otherwise keep the
freeze. Hard read errors keep the existing behaviour (no churn on
unreachable files such as //anon).

The new logic lives in two small helpers (check_jitdump_changed,
read_jitdump) so get_or_insert_jitdump remains a short orchestrator.

Cost: at 99 Hz CPU sampling on up to ~20 cores the worst case is
~30 stat() calls per second profiler-wide, and reparses only happen
when the file actually grew.

Adds three unit tests covering: no growth -> no reread, growth ->
reread + resolve, unreachable path -> no churn.

Co-authored-by: erwan.viollet <erwan.viollet@datadoghq.com>