perf(index): group multi-extension languages in the file-type table

[bis3946]

src/semble/index/files.py maps 347 file extensions to 265 languages. 61 of those
languages cover more than one extension, so their names are repeated 82 times
across the table. Since these are module-level string literals, every duplicate is
interned at import — this module is imported by the indexer, the cache layer and
the CLI, and semble typically runs as a long-lived MCP server.

This splits the table in two:

• _LANGUAGE_EXTENSIONS — the 61 languages reached by 2+ extensions, each named
  once next to its extensions.
• _SINGLE_EXTENSION_LANGUAGE — the 204 extensions whose language is unique,
  unchanged and in the original row order.

_EXTENSION_TO_LANGUAGE is then built from the two. It is still a dict[str, str]
with the same 347 entries; only the source of the data changed.

Why group these at all: the grouped map is also the honest answer to "which
languages have several extensions?" — currently that is spread over 143 rows.

Verification

• _EXTENSION_TO_LANGUAGE compared entry-by-entry against main: 347/347 identical.
• ALL_LANGUAGES, _CODE_LANGUAGES, _DOC_LANGUAGES, _CONFIG_LANGUAGES,
  _DATA_LANGUAGES, _LANGUAGE_TO_EXTENSIONS, _CONTENT_TYPE_LANGUAGES: identical.
• detect_language() probed on 16 filenames; get_extensions() probed on 5
  content-type combinations: identical output.
• No behavioural or API change; tests/test_files.py covers detect_language and
  get_extensions and passes unmodified.

Diff: one file, one module-level table. No tests, no other modules, no new deps.

[stephantul]

Doesn't follow the contribution guidelines, closing.