BUG: pt.roll crashes JAX backend JIT compilation due to dynamic slicing

Describe the issue:

When using pt.roll and compiling the graph with the JAX backend (mode="JAX"), compilation fails.

This happens because roll in pytensor/tensor/basic.py is currently implemented as a macro rather than a dedicated Op. It calculates the shift dynamically (shift = shift % _x.shape[axis]), which turns the shift into a symbolic variable. When it subsequently executes the Subtensor slices, JAX encounters dynamic slice indices (JitTracer) instead of static integers and refuses to compile.

Proposed Fix:
To make this compatible with strict JIT backends like JAX (and Numba), we should elevate roll from a macro to a core primitive Op.

If we create a dedicated Roll(Op), the JAX linker can trivially map it to jax.numpy.roll and bypass the dynamic slicing limitation entirely.

I am happy to put together a PR for this if making it a proper Op aligns with the architectural vision.

Reproducable code example:

import pytensor
import pytensor.tensor as pt

x = pt.dmatrix('x')
out = pt.roll(x, shift=2, axis=0)

# Compiles and runs fine in C
f_c = pytensor.function([x], out)

# Crashes in JAX
f_jax = pytensor.function([x], out, mode="JAX")

Error message:

TypeError: Slice entries must be static integers. Got slice(None, JitTracer(int64[]), None) at position 0
Apply node that caused the error: Join(0, Subtensor{start:}.0, Subtensor{:stop}.0)
Toposort index: 5
Inputs types: [TensorType(int8, shape=()), TensorType(float64, shape=(None, None)), TensorType(float64, shape=(None, None))]

PyTensor version information:

• PyTensor Version: main branch
• Python Version: 3.14
• Operating system: Linux (Ubuntu)
• How did you install pytensor: conda environment (pytensor-dev)

Context for the issue:

Any computational graph, model, or downstream library relying on pt.roll currently cannot be compiled using the JAX backend.

Moving this to a primitive Op not only unblocks JAX/Numba compilation but will also clean up the C backend execution by directly mapping to numpy.roll rather than executing split-and-join operations.

Start in pytensor/tensor/basic.py at the current roll macro and reproduce the failure with the provided JAX-mode function. Trace how its dynamic slicing reaches the JAX linker; done means pt.roll compiles under JAX while preserving the existing C-backend behavior and supporting the proposed primitive-Op design.