fix(kernels): skip non-batch buffers in Kernel.__getitem__ and expand_batch

[umi008]

Closes #2591

Root cause

Kernel.__getitem__ and Kernel.expand_batch apply batch indices to all registered buffers, including non-batch metadata like active_dims. When active_dims has fewer elements than the batch size (e.g. shape (1,) with batch_shape=(2,)), indexing with kernel(x)[1,:,:] raises IndexError: index 1 is out of bounds. When the shapes happen to align, the buffer is silently corrupted to a scalar.

Fix

Add a shape guard in both methods that skips buffers whose leading dimensions do not match the kernel's batch shape. Buffers that genuinely carry the batch dimensions (e.g. parameters, randn_weights) continue to be indexed/expanded exactly as before.

for buffr_name, buffr in self.named_buffers(recurse=False):
    batch_ndim = len(self.batch_shape)
    if batch_ndim > 0 and buffr.shape[:batch_ndim] != self.batch_shape:
        continue
    ...

Changes

File	Change
gpytorch/kernels/kernel.py	Add active_dims/non-batch-buffer guard in __getitem__ and expand_batch
test/lazy/test_lazy_evaluated_kernel_tensor.py	10 regression tests covering scalar index, slice, ScaleKernel nesting, semantic correctness, active_dims=None, direct kernel indexing, and expand_batch scenarios

Test plan

• 10 new regression tests pass (TDD: each was failing before the fix)
• Full test/lazy suite (204 tests) passes
• Full test/kernels suite (466 tests) passes
• pre-commit run --files gpytorch/kernels/kernel.py test/lazy/test_lazy_evaluated_kernel_tensor.py passes (all hooks, including ufmt, pyupgrade, require-ascii)