feat(compile): add --trace/--focus debugging flags

[proggeramlug]

What

Adds --trace/--focus debug flags that consolidate the scattered
compiler-introspection knobs (--print-hir, PERRY_SAVE_LL,
PERRY_LLVM_KEEP_IR) into one CLI surface, plus a focus filter for
localizing "compiled to the wrong thing" bugs — the dominant Perry debug
workflow.

• --trace <stages> — comma-separated hir, llvm, or all.
  • hir: post-transform HIR dump (same data as --print-hir).
  • llvm: writes per-module .ll into .perry-trace/llvm/ by promoting
    the PERRY_SAVE_LL/PERRY_LLVM_KEEP_IR env vars, and forces
    PERRY_NO_CACHE so the per-module object cache can't skip codegen and
    leave the trace dir empty (the same cache-bypass --verify-native-regions
    already does).
• --focus <name> — substring filter on --trace hir: only matching
  functions, class methods, and classes are printed, and the
  imports/exports/init sections are suppressed. One function's lowered body
  is ~40 lines instead of buried in a 10k-line full-module dump. Implies
  hir when no stage is given.

--print-hir keeps its exact historical full-dump behavior (it maps to
focus = None).

Why

When a TypeScript program compiles to the wrong result, ~30% of debug time
is just localizing which stage corrupted a value. The pipeline has clean
stage boundaries but the introspection for them was a scattered set of
undocumented env vars. This makes "dump stage X, focused on function Y" a
one-liner.

Test evidence

$ perry compile demo.ts --trace hir --focus parseRow
=== HIR trace (focus: "parseRow"): demo.ts ===
Functions: 1
  - parseRow (params: 1, ...)
      [0] Let { id: 2, name: "n", ... PropertyGet { ... "length" } }
      [1] Return(Some(Binary { op: Mul, left: LocalGet(2), right: Integer(2) }))
Classes: 1
  - Parser (...)
    - parseRowImpl (params: 1, ...)
        [0] Return(Some(PropertyGet { object: LocalGet(5), property: "length" }))

Verified manually:

• --trace hir --focus NAME → focused dump (functions + matching class methods)
• --focus NAME alone → implies hir
• --trace llvm → .perry-trace/llvm/*.ll written (cache-bypass confirmed)
• --trace all → full HIR + llvm
• --print-hir → unchanged full dump (imports/exports/init present)

cargo fmt --all -- --check clean; new code clippy-clean; cargo build -p perry --tests clean.

Scope / follow-ups (deliberately out of this slice)

• --focus on the LLVM stage (extract just the matching defines post-codegen)
• An ast stage (pre-lowering SWC dump)
• --timings (per-stage wall time)

Per worktree-PR convention this PR omits the version bump + CHANGELOG entry;
fold those in at merge.