feat: improve profile grouping output

[piotrski]

Summary

• improve profiling output for repeated component names by adding grouped profile slow / profile rerenders output
• add parent-path context and source context to make repeated rows easier to act on
• harden metadata collection so profiling reads no longer depend on broad eager inspection or unstable overlapping inspect requests

What changed

• preserve source metadata from inspected elements and normalize it for reporting
• add component path helpers in the tree layer
• group repeated profiling rows only when display name and normalized source metadata match
• keep profile report instance-specific, but show path and src when available
• switch inspect requests to unique request ids and add inspection caching
• remove eager profile-start warmup and full-session metadata fan-out
• enrich only the components needed for profile report, profile slow, and profile rerenders
• keep --limit behavior stable after grouping by enriching a bounded candidate window and trimming at format time

Why

This addresses feedback from #54 about profiling output being hard to interpret when many repeated rows appear, especially Context.Provider-style entries with no useful grouping or hierarchy context.

Limitation

This PR uses the best identity the current backend exposes: normalized source metadata from inspectElement.

That means grouping is conservative and practical, but it is not a true runtime implementation id. We are doing the best we can without backend support for a stable component implementation identifier.

Validation

• pnpm --filter agent-react-devtools test
• pnpm --filter agent-react-devtools typecheck