perf(swc_ts_fast_strip): remove redundant token sorting after parser capture

[kdy1]

Summary

tokens.sort_by_key(|t| t.span) looks redundant because captured tokens are already kept in source order.

Evidence

• Sorting is done in strip mode: crates/swc_ts_fast_strip/src/lib.rs:303.
• Sorting is also done in transform mode when module checking is enabled: crates/swc_ts_fast_strip/src/lib.rs:418.
• Capturing::capture already preserves monotonic order by removing backtracked tokens before pushing (crates/swc_ecma_parser/src/lexer/capturing.rs:43-58).

Why this is a performance risk

Sorting all tokens adds O(n log n) work per call, even though the vector appears to already satisfy the ordering invariant required by binary search lookups.

Suggested direction

• Remove the unconditional sort.
• Keep a debug-only invariant check (for safety), for example using windows(2) assertions on span.lo ordering.
• If any edge case violates ordering, handle it conditionally instead of always sorting.

Benchmark idea

Measure total strip-only runtime and transform runtime (with deprecated module check enabled) before/after removing sort.