cuda: run streaming selected load for single-token ffn batch encodes

[giannisanni]

Fixes the concrete failure behind the --mtp + --ssd-streaming guard on CUDA (context: #495).

metal_graph_cuda_stream_prefill_batch_selected_load() skips when n_tokens <= 1, assuming batch encodes only come from multi-token prefill chunks. The MTP verifier can issue single-position ffn batch encodes; with the selected load skipped, the routed MoE falls through to allocating full resident expert ranges, which by definition cannot fit under SSD streaming:

ds4: CUDA model arena alloc failed for moe_gate (1792.00 MiB chunk): out of memory
ds4: gpu layer 2 ffn batch encode failed
ds4: decode failed: MTP verifier failed

The one-line fix lets n_tokens == 1 take the selected load; the batch machinery already handles a one-token union.

Tested on RTX 4060 Ti 16GB / Linux, Flash q2 streaming from NVMe, with the engine guard bypassed locally for testing:

config	before	after
margin --mtp-draft 2	hard crash (above)	runs, 1.32 t/s
margin --mtp-draft 4	OOM + fallback loops, 0.85 t/s	clean, 0.84 t/s
strict --mtp-draft 2	1.25 t/s, token-identical to greedy	1.18 t/s, still token-identical
non-MTP baseline	2.65 t/s	2.55 t/s (noise)

This does not touch the engine guard itself; whether to relax it (e.g. allow with strict verify) seems like your call. Happy to follow up with that change plus docs if you want it.