Update _PyGCThreadPool doc comment for condvar dispatch

The struct comment still described the old barrier-based dispatch
(mark_barrier/done_barrier broadcast). Those fields were retired by the
FTP dispatch refactor (7d81e3c); update the comment to match what
the struct actually contains: per-worker (wake_mutex, wake_cond) for
dispatch + pool-level (done_mutex, done_cond, workers_done_count) for
completion signalling, with phase_barrier resized per dispatch for use
inside work functions.

Trailing cleanup item from the post-NBS audit. No code change; the
struct fields themselves were removed in the dispatch refactor.

Verified: builds clean and full parallel-GC test suite still green
(GIL: 124 tests OK; FTP previously verified at 177 OK).

Author: SonicField

Include/internal/pycore_gc_ft_parallel.h

@@ -316,13 +316,14 @@ typedef struct {
 // Persistent thread pool for parallel GC
 // Created once on gc.enable_parallel(), destroyed on gc.disable_parallel()
 //
-// Uses barrier synchronization like the GIL-based parallel GC:
-// - mark_barrier: Workers wait here until main signals work ready
-// - done_barrier: All workers (including main as worker 0) wait here when done
-// - phase_barrier: For multi-phase operations (e.g., UPDATE_REFS init/compute)
+// Dispatch uses per-worker condvars (wake_mutex/wake_cond on
+// _PyGCWorkerState) so main can wake only the active adaptive_workers
+// subset; idle workers stay asleep. Worker completions are signalled
+// through the pool's done_mutex/done_cond + workers_done_count.
 //
-// This guarantees correct termination: barriers only release when ALL
-// participants arrive, ensuring no work is in flight.
+// phase_barrier remains for use INSIDE work functions (e.g. update_refs
+// has init/compute phases); it is resized to adaptive_workers per
+// dispatch so it only waits for active participants.
 typedef struct _PyGCThreadPool {
     // Interpreter state - needed to create Python thread states for workers
     PyInterpreterState *interp;