Overflow protection

[jpn--]

The PR adds over- and under-flow protection to the activitysim.core.logit.utils_to_probs function. This error was initially reported by @stefancoe in the PSRC model's mandatory tour scheduling. This PR replaces #679 which got mangled when the ORCA PR was squashed.

• Code developed, all existing tests pass
• PSRC to confirm this resolves their current error.
• Write new unit test to ensure utility shifting works for various corner cases
• Test for performance on full scale model to confirm the computational overhead is acceptably small (tested on PSRC model, no significant change in runtime observed)

The default is now to shift utility values such that the maximum utility in each row is zero, unless allow_zero_probs is set (the overflow protection will spoil this as zero probs are triggered by having an intentional underflow). This constant per-row shift should not fundamentally alter the computed probabilities, but will ensure that an over/underflow does not occur that will create infinite or NaN values. Extremely rare probabilities will round to zero, but by definition they are extremely rare and losing them entirely should not impact the simulation in a measurable fashion, and at least one (and sometimes only one) alternative is guaranteed to have non-zero probability, as long as at least one alternative has a finite utility value. If utility values are certain to be well-behaved and non-extreme, enabling overflow_protection will have no benefit but impose a modest computational overhead cost, but the possibility of encountering a show-stopping underflow accidentally as a corner case probably makes this overhead worthwhile all the time.