btree: make write buffer flush waits SRCU-aware

[Komzpa]

Summary

btree_write_buffer_flush_seq() currently drops the transaction SRCU read lock immediately with bch2_trans_unlock_long() before waiting for write-buffer pins to drain.

Switch that wait to the existing SRCU-aware trans_wait_event() pattern:

• use bch2_trans_unlock() for the fast path;
• wait on write_buffer_flush_wait through trans_wait_event();
• let the helper escalate to dropping SRCU only if the wait blocks long enough.

This is a narrow companion to the broader SRCU/blocking-wait discussion in koverstreet/bcachefs#1069. It follows the maintainer-requested shape there without doing another blanket unlock_long() conversion.

Validation

• git diff --check
• make generate_version
• BINDGEN=/home/kom/.cargo/bin/bindgen make -j32 fs/btree/write_buffer.o
• BINDGEN=/home/kom/.cargo/bin/bindgen make -j32 bcachefs

[koverstreet]

Merged — cherry-picked onto master (bdf3476, authorship preserved). The unlock_long→unlock is required for trans_wait_event to have SRCU to manage, and it's safe: the only thing now running under the held SRCU read lock, fetch_wb_keys_from_journal(), is bounded (buf_lock + write-buffer lock + in-memory key copying — no indefinite wait or journal disk IO), and the actual unbounded wait is the SRCU-aware trans_wait_event. Thanks!