fix: make remaining message and compound-prior xp paths JAX-traceable

[Jammy2211]

Overview

A focused prior/message backend audit performed while fixing #1458 found four additional, independently reproducible JAX tracing failures. Each affected method accepts an xp argument but still calls NumPy/SciPy directly, so a traced value is coerced through __array__ and raises jax.errors.TracerArrayConversionError.

This is not covered by the completed JAX-native prior transforms in #1262 or the NumPy/statistical fixes tracked in #1331.

Confirmed failures

On current main, each of these fails under jax.jit:

• GammaMessage.log_partition(xp=jnp)
  • uses scipy.special.gammaln(alpha) and np.log(beta)
• BetaMessage.log_partition(xp=jnp)
  • uses scipy.special.betaln(*self.parameters)
• Log._instance_for_arguments(..., xp=jnp)
  • ignores xp and calls np.log
• Log10._instance_for_arguments(..., xp=jnp)
  • ignores xp and calls np.log10

Minimal shape of the reproducer:

jax.jit(
    lambda x: GammaMessage(x + 1, x + 2).log_partition(xp=jnp)
)(jnp.asarray(2.0))

Equivalent JIT calls fail for the other three paths.

Proposed fix

• Dispatch Gamma/Beta special functions through a JAX-compatible backend when xp is JAX, while preserving SciPy on the NumPy path.
• Use xp.log / xp.log10 in the compound prior modifiers.
• Confirm whether compound priors must forward xp recursively to child prior models; add coverage for nested modifiers/compound expressions if so.
• Add scalar and batched jax.jit regression tests with NumPy/SciPy numerical parity.
• Keep statistical semantics unchanged.

Verification

• Direct tests must fail on current main with TracerArrayConversionError and pass after the fix.
• Run the focused prior/message suites and full PyAutoFit suite.
• Exercise at least one nested compound-prior expression through the real instance_for_arguments(..., xp=jnp) API.

Relationship to #1458

#1458 addresses fixed-array construction and GammaMessage.to_canonical_form. These four failures are separate call paths and should remain a follow-up PR to keep that repair reviewable.

[Jammy2211]

Shipped in #1461 (squash commit fbe9f45deac2f8f4646d61bdf5f4cb75ac950e30).

Validation before merge:

• GitHub Docs: passed
• GitHub Tests: passed
• supported JAX 0.7.0 direct suite: 32 passed
• JAX 0.10.2 direct suite: 32 passed
• focused prior/message suites: 317 passed
• full local PyAutoFit suite: 1701 passed, 4 skipped
• untouched post-Use xp.stack in messages so JAX can trace fixed arrays #1460 main: 12 expected regression failures, 20 passes

The implementation also fixed the earlier Beta/Gamma traced-constructor np.broadcast boundary discovered while exercising the original four backend leaks.