[Benchmark] Add compute_seq_len_sweep_config_with_probe with linear/quadratic scaling support

[shivam2199]

Summary

Refs #1200. Addresses non-linear memory scaling in benchmark sweep config inference.

The existing compute_seq_len_sweep_config inverts memory via max_tokens = usable_bytes / kernel_bytes_per_token, which only holds for linear-scaling kernels. For O(L²) kernels (e.g. benchmark_sparse_multi_token_attention.py), this overestimates capacity by orders of magnitude — the existing workaround there divides by probe_L * probe_L, but the downstream sweep math still treats the result as linear bytes-per-token.

Per discussion on the issue (#1200 (comment)), this PR adds a new helper rather than threading scaling_method through the existing function — 16+ benchmark scripts call estimate_kernel_peak_memory today, and a wider signature change would conflict with in-flight benchmark refactors (#1199, #1180). Linear-scaling callers are unchanged; only quadratic-scaling benchmarks opt in.

What changed

• benchmark/scripts/benchmark_model_configs.py — adds compute_seq_len_sweep_config_with_probe(model_cfg, probe_fn, probe_seq_len, probe_batch_size=1, scaling_method="linear" | "quadratic", ...). Internalizes the probe call + inversion; reuses estimate_kernel_peak_memory for the measurement.
• benchmark/scripts/benchmark_sparse_multi_token_attention.py — switches the token_length sweep mode to the new helper with scaling_method="quadratic", dropping the manual peak_bytes // (probe_L * probe_L) workaround.

estimate_kernel_peak_memory and compute_seq_len_sweep_config are untouched.

Validation

Hardware: A10G 24GB (g5.xlarge).

Synthetic O(L²) probe (B=2, L=2048, allocates B * L * L floats) using LLAMA_3_8B config and max_seq_len=2**20 to bypass the model cap so the raw inversion is visible:

quadratic: SeqLenSweepConfig(batch_size=2, seq_len=8192)
linear:    SeqLenSweepConfig(batch_size=2, seq_len=65536)

The 8× gap (≈17× before snap-to-power-of-2) demonstrates the inversion difference: linear claims a sweep at L=65536 fits, when in reality L² at that size would require multiple TBs. quadratic lands at a realistic L=8192. This matches the issue's premise — for non-linear-scaling kernels, the existing inversion overestimates capacity and would OOM at the predicted boundary.

Testing Done

• Synthetic O(L²) sanity check on A10G — confirms quadratic predicts L=8192 vs linear predicts L=65536 for the same probe (8× separation, scales as expected).
• benchmark_sparse_multi_token_attention.py imports + helper resolution verified locally.
• Full sparse-attention end-to-end sweep on A10G (deferred — synthetic test already isolates the inversion math from kernel-specific noise).

cc @Tcc0403