perf(global): cache prefix records, parallelize conda-meta parsing, fast trampoline check

[Hofer-Julian]

pixi global commands re-parsed every environment's conda-meta/*.json —
tens of MBs of JSON for Python-sized environments — many times per
invocation: the sync check, exposing executables, shortcuts and
completions each triggered a full parse, plus one parse per direct
dependency when looking up designated packages.

• Cache the parsed prefix records per environment on Project, shared
  across clones and invalidated whenever the prefix is modified
  (install, environment removal, pruning, and menuinst rewrites).
• Parse conda-meta records in parallel (rayon via
  Prefix::find_installed_packages) in find_package_records instead of
  reading them sequentially one file at a time, and reuse the project
  cache in the list commands so 'pixi global list' parses each
  environment once instead of twice.
• Answer Trampoline::is_trampoline from file metadata in the common
  case: verify the shared trampoline binary against the embedded one
  once per process, then treat hard links to it as trampolines instead
  of reading and comparing the full ~1MB binary for every exposed
  executable on every sync.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01UZeFV9B6ECmc43Rppj8G1U