perf: checking the syntax contains/cluster list is slow

[h-east]

Problem:  Deciding whether a group is in a "contains"/cluster list scans
          the list and expands clusters on every check, which is slow for
          syntaxes with large lists (e.g. plugins such as netrw).
Solution: Resolve each list once into a sorted, cluster-expanded set of
          group IDs and use a binary search; cache it per syntax block and
          drop the cache when syntax definitions change.

[h-east]

Measurement results

Build: ./configure default CFLAGS (-g -O2), no profiling.

Method: clean A/B on the same binary using test_override('syn_idlist_cache',
{0|1}) to enable/disable the cache; force full-buffer highlighting by calling
synID() on every line; measure wall time with reltime(), median of 5 runs.

file (filetype)	lines	cache off	cache on	change
netrw.vim, Vim plugin	9,717	~5.5 s	~3.7 s	~33% faster
src/evalfunc.c, C	12,919	~0.41 s	~0.35 s	~14% faster
big.c, C (concatenated)	99,192	~2.7 s	~2.5 s	~10% faster

in_id_list() self time drops from ~10.7% to ~6.6% (perf). The gain is largest
for syntaxes built around "contains"/cluster lists (real-world plugins such as
netrw); for keyword-heavy or small buffers it is smaller.

Correctness: byte-for-byte identical synID() output with the cache on and off
across C, C++, Vim script, Python, Ruby, Lua, JavaScript, shell, HTML and CSS
(millions of cells), including netrw.vim and the Vim syntax file.

Tests: test_syntax (incl. a new Test_syntax_idlist_cache_unchanged differential
test), test_highlight and test_functions all pass.

Maintainability

• The cache is a speed optimization only. test_override('syn_idlist_cache', 1)
  disables it at runtime (same mechanism as nfa_fail), and
  Test_syntax_idlist_cache_unchanged asserts the highlighting is identical with
  the cache on and off, so a future bug surfaces as a CI failure rather than
  silently wrong highlighting.
• Lists that cannot be represented as a plain set (a nested
  ALLBUT/TOP/CONTAINED marker) fall back to the original linear scan.
• The cache is dropped in syn_stack_free_block(), which runs on every syntax
  definition change, so it also covers list pointer reuse.

Relationship to the lead-byte prefilter PR (vim#20371)

This is a separate, complementary optimization: the prefilter speeds up
regexp-pattern-heavy syntaxes (C/C++), while this speeds up
contains/cluster-heavy ones (real Vim plugins). As currently written the work
is stacked on the prefilter branch (it reuses the s:SynDumpBuffer test helper),
so it is most naturally a follow-up/stacked PR; making it fully independent of
vim#20371 would require duplicating that helper.