feat: add roofline analysis (RooflineAnalyzer, measure_peaks)

[nathanhubens]

Summary

Adds roofline-model measurement primitives to fasterbench. Per the measurement-only contract, this module exposes numbers and a plot; compression decisions live in fasterrecipes.

New public API

• measure_peaks(device, *, dtype, matmul_size, bandwidth_mb, warmup, steps, allow_tf32, cache) — empirical probe of achievable peak FLOPs/s (via large matmul) and streaming bandwidth (via cache-defeating memcpy). Pins TF32 off by default for honest fp32 peaks. Records tf32_enabled and cudnn_benchmark flags in the result.
• HardwarePeaks — dataclass with peak_flops, peak_bandwidth, ridge_point, device, dtype, and the pinned flags.
• RooflineAnalyzer(model, sample, peaks=None) — per-layer profiler with .profile(device, warmup, steps), .summary(top), .plot(title), .results. Single-pass forward hooks measure FLOPs (analytically for Conv{1,2,3}d, ConvTranspose{1,2,3}d, Linear), bytes moved (weights + input + output per Williams 2009), and wall time. Layers outside those types get bound="undefined" with a warning — no silent zeros.
• RooflinePoint — per-layer result with flops, bytes_moved, time_s, arithmetic_intensity, achieved_gflops, bound ("memory"/"compute"/"undefined"), utilization_pct.
• clear_peaks_cache() — explicit cache reset.

Deviation from the initial plan (worth flagging)

The plan called for a two-pass profile using torchprofile for FLOPs and a separate timing pass. The installed torchprofile build had a broken handler dispatch that silently returned zero FLOPs per layer, so the implementation switched to analytical FLOPs for Conv/Linear in a single pass. Trade-offs: exact math, no cross-pass drift, no external profiler dependency — at the cost of narrower layer-type coverage. Other layer types fall into the already-planned "undefined" bucket, so this was an architecturally clean forced deviation.

Tests

• 4 inline unit tests: basic peaks probe, cache identity, hand-computed Conv2d (flops=36864, bytes=4224, AI≈8.73), Linear stack with synthetic peaks.
• 2 #|slow integration tests: ResNet-18 CPU with synthetic peaks, CUDA smoke test guarded by is_available().
• Tutorial at nbs/tutorials/roofline.ipynb showing hardware peaks, per-layer profiling on ResNet-18, and AI shift when input resolution changes from 224×224 to 512×512.

Architectural boundary

• No imports from fasterai / fasterlatency / fasterrecipes.
• No bottlenecks() / suggest() / recommendation helpers.
• Tutorial avoids prescriptive language; includes a callout directing users to fasterrecipes for compression decisions.

Test plan

• nbdev-test --path nbs/analysis/roofline.ipynb passes
• nbdev-test --path nbs/analysis/roofline.ipynb --flags slow passes
• nbdev-test (full suite) passes
• Tutorial executes end-to-end (jupyter nbconvert --execute)
• No prescriptive language ("consider", "recommend", "try", "should compress") in notebook or tutorial
• No cross-package imports (only fasterbench.core / fasterbench.profiling + torch / numpy / plotly)
• nbdev-clean leaves a clean checkout
• CI passes on feature/roofline