Allow concurrent parsing of the same grammar during indexing #127

Summary

PR #126 (#115) parallelizes per-file indexing, but chunk() in crates/csp/src/chunking/core.rs holds a per-grammar mutex around Parser::new, set_language, and parse. Files of the same language therefore still parse one at a time. On a repository dominated by one language, the speedup comes only from reading, hashing, merging, and tokenizing.

The lock exists because the language pack serializes all parsing behind a global PARSE_LOCK ("some third-party grammar scanner code keeps process-global mutable state"), and upstream semble sidesteps the issue with worker processes. The per-grammar lock is already more concurrent than the language pack's own default.

Proposal

• Identify which grammars' external scanners actually keep process-global state, and keep an allowlist of grammars that are safe to parse concurrently (or a denylist of unsafe ones).
• Alternatively, keep a per-thread parser pool for grammars known to be safe.
• Add a single-language benchmark case (for example 300 .ts files) next to the existing multi-language index_build bench, so the effect is measurable.

Context

• Raised in review on perf(index): parallelize per-file read, hash, chunk, and tokenize (#115) #126 (cubic).
• The trade-off and this follow-up are noted in the perf(index): parallelize per-file read, hash, chunk, and tokenize (#115) #126 PR description.