Use hosted NCU profiling by default without provider setup

[msaroufim]

popcorn submit submission.py --profile now submits through the authenticated GPU Mode API and downloads NCU reports. Users need only normal Popcorn registration: no Modal SDK, provider account, credentials, or profiling URL. B200 is the default when no GPU flag or directive is supplied.

--profile-brev explicitly selects the existing Brev service. Errors never switch providers, and --local evaluation remains separate. Both profiling paths accept benchmark selection and kernel-name/launch-count filters. Hosted results include .ncu-rep, text/CSV detail exports, and a manifest with selected benchmark specs, capture options, system information, and a configuration fingerprint. Failed or empty captures return errors.

Deployment dependency: merge and deploy gpu-mode/kernelbot#530 (API and GPU runners) before releasing this CLI. It adds the /profile/{leaderboard}/{gpu} endpoint and pins the working server-side NCU version. The new-problem guide is gpu-mode/reference-kernels#172. No production deployment is included here.

Validation:

• 59 Rust tests passed; cargo fmt --check, cargo clippy --all-targets -- -D warnings, and release build passed.
• Authenticated HTTP test verifies the Popcorn header and multipart capture options, with SSE errors propagated.
• Isolated GPU integration exercised the real API/backend/runner with a fake database and ephemeral B200 launcher; the release CLI ran with an empty PATH and no MODAL_* values, then saved and extracted a 7,593,257-byte report and both exports.
• Fixture: problems/linalg/qr_v2, benchmark 0 (batch=20, n=32, cond=1, seed=43214), reference-kernels 51e22db671d36c1c76091c43c36a44546ba324a1, regex:geqr2, demangled names, launch count 1. NCU 2025.2.1 collected 39 passes and valid occupancy, throughput, cache, and instruction counters. Six ctc__ metrics were unavailable. Other tasks require NCU-compatible evaluators and their own GPU validation.

docs/modal-ncu-validation.md records the integration boundary and the NaN-counter issue observed with NCU 2026.2, which motivated the backend's version pin.