fix: release flights after compute panic

[agkloop]

Why

A panic from a foreground GetOrCompute computation left its single-flight entry unfinished. Later calls for the same key could block forever.

What changed

• Preserve and re-panic the original value for the computing caller and same-flight followers.
• Remove the failed flight before releasing waiters.
• Cache no value or error after a panic, so a later call can retry normally.
• Keep background-refresh panic behavior unchanged.
• Document the behavior and add regression coverage for panic propagation and retry.

Validation

• go test ./... -count=1
• go test ./... -race -count=1
• Redis adapter tests and race tests
• Focused panic tests repeated 100 times under the race detector
• go vet ./...
• docs/skill sync check
• benchmark smoke: hot hits remain zero-allocation