Heal candidates in parallel in the pytest plugin

[Johin2]

Problem

The pytest plugin heals candidates one at a time:

https://github.com/Johin2/self-heal/blob/main/src/self_heal/pytest_plugin.py#L127-L131

Each candidate is an independent, network-bound LLM round-trip, often several attempts deep. A suite with eight @pytest.mark.heal failures pays all eight sequentially, so a run that could take one round-trip's wall clock takes eight. This is the difference between --heal being something you run on every red suite and something you run when you have a coffee break.

Proposal

Run candidates concurrently with a bounded worker pool.

• --heal-workers N (default something conservative like 4, 1 restores today's behavior).
• Use threads: the work is I/O-bound on the provider, and the existing async proposer paths can be driven from a thread pool without restructuring the loop.
• Keep terminal output deterministic by buffering each candidate's report and flushing in the original candidate order, not completion order.

Acceptance criteria

• Candidates are healed concurrently up to the worker limit.
• Reported output ordering is identical to the sequential run.
• --heal-apply writes are serialized; two candidates in the same file must not race on the patch.
• A failure in one candidate cannot abort the others.
• --heal-workers 1 is byte-identical to current behavior.

Notes

The apply-path serialization matters more than it looks: with #58 landed, two failing methods in one class are two candidates pointing at the same file.