Add specialized float32 numba dispatches for transcendental functions

Description

tanh, cos, sin, exp, and erfc admit polynomial approximations that offer nice speedups over what we currently have. Numbers from a very quick claude session:

  float32, 12.6M elements, single-threaded, min of 15                                                       
                                                                                                            
  ┌──────┬──────────────────┬─────────────────┬─────────┬─────────┬─────────────┐                           
  │  op  │ now (numba libm) │ with polynomial │ speedup │  saved  │    error    │                           
  ├──────┼──────────────────┼─────────────────┼─────────┼─────────┼─────────────┤                           
  │ cos  │ 67.50 ms         │ 7.79 ms         │ 8.66x   │ 59.7 ms │ 5.5 ULP abs │                           
  ├──────┼──────────────────┼─────────────────┼─────────┼─────────┼─────────────┤                           
  │ sin  │ 61.25 ms         │ 7.21 ms         │ 8.50x   │ 54.0 ms │ 5.9 ULP abs │                           
  ├──────┼──────────────────┼─────────────────┼─────────┼─────────┼─────────────┤                           
  │ erfc │ 102.58 ms        │ 45.42 ms        │ 2.26x   │ 57.2 ms │ 4.0 ULP rel │                           
  ├──────┼──────────────────┼─────────────────┼─────────┼─────────┼─────────────┤                           
  │ tanh │ 48.67 ms         │ 12.61 ms        │ 3.86x   │ 36.1 ms │ 2.7 ULP rel │                           
  ├──────┼──────────────────┼─────────────────┼─────────┼─────────┼─────────────┤                           
  │ exp  │ 17.70 ms         │ 5.69 ms         │ 3.23x   │ 12.0 ms │ 2.3 ULP rel │                           
  └──────┴──────────────────┴─────────────────┴─────────┴─────────┴─────────────┘

These functions show up in e.g. machine learning where we want to do float32, so it might be nice to implement these. On the other hand, they are less accurate, especially erfc/exp in the tails. I could do them as specialty dispatches in pytensor-ml, but I thought they might be more broadly interesting hence this issue first.

Start by locating the existing Numba/libm dispatch path for tanh, cos, sin, exp, and erfc and review how float32 inputs are handled. Define acceptable accuracy, tail behavior, and benchmark expectations for each polynomial implementation; done means the specialized dispatches are integrated and their speed and error are validated.