feat: add --profile flag for LLM-readable CUDA kernel summary

[Bortlesboat]

Problem

Addresses #118. The AI agent has no visibility into which CUDA kernels are the actual hardware bottlenecks. It can only infer from step time and MFU%, which are too coarse to guide targeted optimization (e.g., "is attention or MLP the bottleneck?").

Solution

Add a --profile flag to train.py that:

1. Runs torch.profiler over a small number of warmup steps (2 warmup + 5 active)
2. Prints a Markdown table of the top CUDA kernels sorted by self-time
3. Exits cleanly — does not consume any of TIME_BUDGET

Usage:

uv run train.py --profile

Example output:

## Top 15 CUDA kernels (5 profiled steps)

| Rank | Kernel | Self CUDA (ms) | % of Total | Calls |
|------|--------|---------------|------------|-------|
| 1 | `flash_attn_varlen_fwd` | 1842.31 | 38.2% | 35 |
| 2 | `volta_sgemm_128x64_nn` | 1204.17 | 24.9% | 70 |
| 3 | `elementwise_kernel` | 412.88 | 8.6% | 140 |
...

_Total CUDA time: 4.823s over 5 steps_

This lets the agent immediately see, for example, that FlashAttention dominates at 38% — and decide whether to explore window size changes vs. MLP width changes.

Changes

• train.py only (the single file agents are allowed to modify)
• +sys, +argparse imports (~2 lines)
• argparse.parse_known_args() block (~5 lines, add_help=False so it doesn't conflict with uv run)
• run_profiler() function (~35 lines)
• Guard before training loop: if PROFILE_MODE: run_profiler(...); sys.exit(0) (~4 lines)

No changes to prepare.py, program.md, or any other file. Normal training runs are completely unaffected.