ENH: general framework for non-overlapping switch logp transforms (monotone branches)

[eclipse1605]

Description

PR #7995 adds logp support for a restricted non-overlapping switch transform (leaky-ReLU pattern) by gating between branch logps evaluated at the observed value. The next step is to track a broader framework to support switch(cond(x), f_true(x), f_false(x)) where:

• x is a single continuous measurable RV
• both branches are measurable (PyMC already knows logp(f_true(x), value) and `logp(f_false(x), value)).
• The branch maps are monotone on their respective domains and have non-overlapping images, so value deterministically implies which branch was taken

The goal is to extend beyond pure scaling (piecewise linear) to a curated set of monotone transforms where we can safely infer a predicate value_implies_true_branch(value) plus runtime constraints that make the inference valid. The switch logp should remain a thin dispatcher: logp = switch(value_implies_true_branch(value), logp(true_branch, value), logp(false_branch, value))

Examples of target patterns:

• switch(x > 1, x**3, x) (true branch implies value > 1)
• switch(x > 0, exp(x) - 1, x) (true branch implies value > 0)
• switch(x > -1, x, exp(x+1) - 2) (true branch implies value > -1)