feat(pr-review): add optional max_turns input to override turn budget

[Troublor]

Summary

Makes the per-invocation turn budget of the claude-pr-review composite action configurable per consumer repo.

Today max_turns is computed entirely inside the action's compose step (12 fast / 44 standard / 56 deep) and is not exposed as an input, so a consuming repo can't raise or lower the analysis ceiling without editing the centralized action (which ships to every consumer at once).

This adds an optional max_turns input:

• Empty (default) → behavior is byte-for-byte unchanged: the existing model-tier / review-depth logic still picks 12 / 44 / 56, and the retry still gets 1.5×.
• Positive integer → overrides the main analysis pass; the retry keeps its 1.5× headroom relative to the chosen value ((n*3+1)/2).
• Invalid value (non-integer or 0) → fails the compose step loudly with a clear ::error:: rather than silently disabling the turn cap.

Backward-compatible: existing consumers pass nothing and are unaffected.

Changes

• action.yml: new max_turns input; wired into the compose step as MAX_TURNS_OVERRIDE, applied after the tier/depth defaults with up-front validation.
• README.md: document max_turns, and fix stale turn-budget numbers — the README stated 36 for the strong tier in one place and 24/40 for standard/deep in another, both of which disagreed with the action's actual 44/56.

Verification

• YAML parses (yq lists the new input in the right position); actionlint clean (its "jobs/on missing" notes are the usual false positives from treating a composite action.yml as a workflow).
• Simulated the compose turn-budget logic across all branches — defaults reproduce today's exact values (44/66, 12/18, 56/84), overrides win over both fast tier and deep depth, and abc/0 are rejected:

tier	depth	max_turns	result (main / retry)
strong	standard	(empty)	44 / 66
fast	standard	(empty)	12 / 18
strong	deep	(empty)	56 / 84
strong	standard	100	100 / 150
fast	standard	80	80 / 120
strong	deep	30	30 / 45
strong	standard	abc	❌ step fails
strong	standard	0	❌ step fails

• The existing actions-test.yml Python unittest suite is unaffected (the change is in the compose shell step, not review_pipeline.py).

🤖 Generated with Claude Code