feat(fit): progress logging + skip_candidates knob

[adrida]

tracer.fit() previously produced no output during potentially-long candidate sweeps and pipeline-family construction. On multi-class high-dim embedding inputs (e.g. 22-class × 1024-dim BGE-M3) a single stuck candidate (GradientBoosting) silently stretched a sweep from ~13s to ~20 minutes with no signal at all.

This wires up default per-candidate + per-stage progress to stderr and exposes a configuration knob to drop known-slow candidates without monkey-patching.

Changes

• FitConfig.verbose: bool = True — gates a default stderr logger
• FitConfig.skip_candidates: tuple[str, ...] = () — drop candidates by name
• surrogate._candidates(n, skip=()) — new parameter
• surrogate.search_best_surrogate now accepts skip= and tolerates both the legacy on_candidate(name, f1) signature and a new 3-arg variant on_candidate(name, f1, elapsed_seconds). Callback exceptions are warned and swallowed so a buggy logger can't abort a fit.
• fit.pipeline.build_global / build_l2d / build_rsb / _build_accepting_stage / fit_frontier all accept log= and skip=, emit stage-start and stage-end heartbeats, and forward skip= to the surrogate sweep.
• api.fit constructs the default stderr logger when verbose=True and threads log/skip through to fit_frontier.
• CLI demo's monkey-patched _candidates / search_best_surrogate are updated to accept the new parameters transparently.

Tests (tests/test_fit.py)

• callback invoked once per candidate under both legacy and extended signatures
• raising callback does not derail the sweep; warning is emitted
• skip_candidates removes named models from the factory
• verbose=True emits [tracer.fit +Xs] heartbeats to stderr
• verbose=False remains silent

All 21 tests pass locally.