Split the layers CI group and cut its redundant tests

[jessegrabowski]

The layers group in the run_tests.yml test matrix is now three groups, layers, recurrent and convolution. The old group took 18 to 24 minutes per run. tests/test_recurrent.py and tests/test_conv.py alone accounted for 11 of its 20 minutes of test time.

The six test files in those groups went from 245 tests to 147, and their cold run on 8 workers went from 54 s to 20 s. Most of the removed tests repeated a stronger reference comparison, or only asserted that a training loss fell. _extract_patches, _scatter_patches and _window_indices are removed from pytensor_ml/layers/conv.py. They had no callers outside the tests that used them as a reference implementation.

Coverage was added at the same time:

• the ConvLayer bias gradient
• Im2Col.perform and Col2Im.perform
• pooling dilation
• per-axis conv padding

Before this, none of these had a test that failed when it broke.

Recurrent._masked_step now zeroes masked inputs before the step. NaN in the padded positions made every parameter gradient NaN, even though the forward outputs stayed finite.

Upsample2D(mode="bilinear") now returns its input's dtype. It built the interpolation weights at floatX, so a float32 input came back as float64 under the default config.

Closes #166

---

📚 Documentation preview 📚: https://pytensor-ml--167.org.readthedocs.build/en/167/