X_logit_{rng, lpdf, lupdf}(theta) should allow -infinite values in theta

The safety checks in the current implementations of categorical_logit_rng and multinomial_logit_rng are too strict. The doc says categorical(softmax(x)) is the same as categorical_logit(x), but that's not the case because softmax allows negative infinite inputs.

There are two things to fix:

1. Remove the bounds checks in the X_logit_rng functions to allow -infinity.
2. [Optional] Allow a single +infinity value to produce a deterministic distribution on that value. Multiple +infinity values could either throw or you could make it uniform among the +infinity positions. I'm not sure which mesas more sense.
3. Remove the bounds check in X_logit_lpdf and X_logit_lupdf when the argument is a data variable.
4. Throw an exception or make the result uniform if all inputs are negative infinity. Again I'm not sure which makes sense.

For 3, we still need to flag cases where the argument is an autodiff variable because the infinite values will wreak havoc with derivatives.

Start by locating the implementations of categorical_logit_rng, multinomial_logit_rng, and the X_logit_lpdf/lupdf functions, then inspect their current bounds checks and how they distinguish data from autodiff arguments. Determine and document behavior for negative infinity, positive infinity, and all-negative-infinity inputs before changing the checks. Done means data arguments accept the supported infinite values, autodiff arguments remain protected, and tests cover the selected edge-case behavior.