feat: compiler-phase timing flamegraphs from tracing spans

[TomAFrench]

Description

Problem

We have no easy way to see where nargo compile spends its time. The compiler already emits tracing spans as JSON logs (via NARGO_LOG_DIR + NOIR_LOG), but nothing consumes them, and some key phases (individual SSA passes, parsing) had no spans at all.

Summary

Adds a compiler-phase profiling workflow:

just profile-compiler test_programs/benchmarks/semaphore_depth_10

produces, from the span logs of a single nargo compile run:

• compiler-flamegraph.svg — an aggregated, self-time-weighted flamegraph of compiler phases (inferno).
• compiler-trace.json — a per-invocation timeline that opens in Perfetto.

Changes:

• Instrumentation: a ssa_pass{pass=...} span around every SSA pass in SsaBuilder::try_run_pass, and a span on parse_program (lexing is lazily interleaved with parsing, so there is a single combined span).
• Post-processor: a new internal dev tool, tooling/compiler_profiler (noir-compiler-profiler), that streams the JSON span logs, reconstructs the call tree from close events (which carry the authoritative ancestor chain plus time.busy/time.idle), and emits both artifacts. It is deliberately not part of the user-facing noir-profiler, which profiles Noir programs rather than the compiler.
• Untracked time is explicit: un-instrumented time inside a phase shows as a (self) frame, wall-clock time outside any root span as (untracked), and the tool reports what fraction of wall clock the flamegraph accounts for, warning on anomalies (unclosed spans, children exceeding their parent).
• Cross-thread spans (parallel workspace parsing, compile_program itself) appear as detached roots in tracing; they are re-parented by time containment.
• just profile-compiler recipe + tooling/compiler_profiler/README.md documenting usage, span filters, and how to read the output.

Per-pass durations in the flamegraph match nargo compile --benchmark-codegen to the millisecond across all 74 pipeline steps.

The default span filter (trace,noirc_frontend::elaborator=info) keeps logs small and timings accurate; full NOIR_LOG=trace additionally captures the elaborator's per-expression spans for a deep frontend breakdown at the cost of multi-GB logs and significant observer distortion.

Additional Context

The flamegraph for semaphore_depth_10 (~2.2s cold compile) breaks down as: ~0.4s parallel parsing, ~0.6s check/elaboration, ~1.1s SSA passes + ACIR gen, ~40ms monomorphization.

Documentation

Check one:

• No documentation needed.
• Documentation included in this PR.
• [For Experimental Features] Documentation to be submitted in a separate PR.

PR Checklist

• I have tested the changes locally.
• I have formatted the changes with Prettier and/or cargo fmt on default settings.