Handle subnormal mixture weights without XLA flush-to-zero

[FlorianPfaff]

Bug

AbstractMixture validates explicit weights on the host, but then determines nonzero components and normalizes using backend comparisons/reductions. On JAX/XLA, representable positive float64 subnormals can be flushed to zero during those operations. A valid mixture such as weights proportional to [2e-311, 1e-311] is therefore rejected as having no nonzero components; the existing backend fallback also sees a zero max and cannot recover the ratio.

This is inconsistent with the robust XLA-subnormal handling already used by AbstractDiracDistribution.

Fix

• determine exact zero/nonzero support from the already validated host representation;
• retain the existing backend normalization path for ordinary weights and the existing square-root scaling path for large finite weights;
• if backend arithmetic loses the weight scale, normalize the finite nonnegative host values after max-scaling and move only the ordinary-sized probabilities back to the active backend;
• tolerate floating-point overflow traps in backend reductions so the stable fallback can run.

Regression

Adds a focused test with two positive float64 subnormal weights in a 2:1 ratio. The regression verifies that neither component is pruned and that the resulting weights are [2/3, 1/3].

The branch is based on current main after #5411 and has no overlap with the open PR queue.