[perf]: replace eval() with direct comparison in check_param

[vTusharr]

Replace eval('np.' + repr(param_dt) + constraint) with a lookup dict of direct numpy comparison operators in check_param(). This eliminates many calls to eval() and repr() per run ; Falls back to eval() for unknown constraint strings to maintain backwards compatibility.

Summary

check_param() previously built ->msg = param_name + f" = {repr(param)}"
at the start of the function. This forced to compute the string and run repr(param) on every call, even when no error occurred.

Change

The message construction is moved into a nested helper:

def _build_msg():
    return param_name + f" = {repr(param)}"

Error paths now call it only when needed:

raise TypeError(_build_msg())

###why?

This uses lazy evaluation while keeping the code DRY . Instead of computing the message on every call or duplicating the string-building code across multiple error sites, it is built only when an exception is raised.

###now
Avoids unnecessary repr(param) and string formatting in the common case where parameters are valid, removing overhead from the hot path and producing a significant speedup.

py-spy flame graph

before

after

Profiling methodology

• Tools: cProfile, py-spy (top + record), flameprof
• profiling across 12 operation categories (fit, gridsearch,
  predict, intervals, sampling, partial dependence, constraints, etc.)

tests 162 passed, 1 skipped, 62 warnings in 30.35s

[dswah]

@vTusharr Cool thanks for the PR.
Can you help me understand if this change improved performance?
I see the flame graphs, but i cant tell if it is faster with your contribution.

[vTusharr]

@dswah Yeah , was kinda busy with exams & all , Old code was calling eval() and building error message strings on all calls, even when validation passed , during a gridsearch that's a lot of calls , so i made it so that only when it fails the string is constructed and replaced eval("np." + repr(param_dt) + constraint) with a dict of pre-built lambdas