JAX linker: Cannot compile Convolve1d without static mode

Problem

The JAX linker cannot compile Blockwise{Convolve1d} without static (known-at-compile-time) shapes.

NotImplementedError: Cannot compile Convolve1d to jax without static mode

MRE

import pytensor.tensor as pt
from pytensor.compile import Mode, function

x = pt.matrix("x")
kernel = pt.vector("k")
# Adstock-like convolution with dynamic shapes
result = pt.xlogx.xtensor.signal.conv1d(x, kernel, mode="full")
fn = function([x, kernel], result, mode=Mode(linker="jax", optimizer="fast_run"))

Context

This prevents the pymc-marketing MMM model from using the JAX linker for MCMC sampling. The adstock transformation uses Blockwise{Convolve1d} with dynamic shapes — the Numba and CVM linkers handle this correctly.

uv pip install jax jaxlib to reproduce.

Start with the supplied Python MRE and trace the JAX linker path for Blockwise{Convolve1d}, focusing on why dynamic shapes are rejected while Numba and CVM support them. Reproduce the failure with the listed JAX installation, then verify that the MRE compiles and runs with dynamic shapes without regressing static-shape behavior.