Vectorize angular (ATS) reductions to fix XLA compile blowup

[jpbrodrick89]

The angular forward model wrapped two functions whose Python list comprehensions were fully unrolled into the jit-traced graph, once per output bin, and then re-traced through value_and_grad:

• thomson_diagnostic.reduce_ATS_to_resunit: ~1024 + 1024 slice+mean+stack ops to bin the wavelength and angular axes.
• irf.add_ATS_IRF: ~2048 + 1024 jnp.convolve ops for the separable 2D instrument-response smoothing.

At angular shapes (npts=2048, 1024 angles) this is ~5000 unrolled iterations baked into a single graph. On CPU the angular value_and_grad failed to finish compiling in 15+ minutes.

Replace both with vectorized equivalents:

• _bin_average: reshape (+ NaN-pad for a ragged final window) and mean, a single op; bit-for-bit equal to the loop including the ragged case.
• add_ATS_IRF: vmap each 1D convolution over the batch axis.

After the change the same angular value_and_grad compiles in ~4.6s on CPU. Outputs are unchanged (1D fits recover identical parameters).

This is done with vmap not scan so both CPU and GPU should benefit equally.

Add tests/test_forward/test_ats_vectorization.py pinning both vectorized forms to the original loop semantics — these run on CPU, unlike the angular integration tests which skip without a GPU.

[almilder]

Looks like a great fix, thanks for adding this!