Compile the coverage loss with the model when compile is set

[davek44]

With compile: true, only the model was compiled; the elementwise poisson_mn loss over B × T × L ran eagerly. This PR compiles the loss too (dynamic=False, one specialization per dataset/batch shape). MLM is excluded.

Evidence (RTX 4090, L 16,384, no graph breaks, 6 graphs — one per shape and grad mode):

case	rel loss	rel grad	eager ms	compiled ms
human (B 4, T 768)	1.1e-7	2.2e-5	10.19	1.79
mouse (B 4, T 384)	0	2.4e-5	5.10	0.90
human, track weights	1.1e-7	3.0e-5	10.19	1.79
human, final batch (B 2)	1.1e-7	2.2e-5	5.09	0.90

5.7× faster than eager, with ~1/5 the transient memory. Each new (dataset, batch shape) pays one compile, like the model itself.

🤖 Generated with Claude Code