gh-114682: Improve performance of module state access in _decimal

[neonene]

• Fix the module state access in ieee_context().
• Optimize get_module_state_by_def(), which *_richcompare() now switch to.
• Optimize find_state_left_or_right().
• Make Context objects carry decimal_state.

• Issue: _decimal: pi benchmark 31-45% slower compared to 3.9 #114682

[neonene]

Around 18% recovery with na+8 and da+32 in the pi test for me, which is no slower than the global state access. About 14% with 8+na and 32+da. FYI,

• Carrying a module state by the Decimal object seems not to be faster even if dec_richcompare() is boosted.

• SignalDictMixin type has non-NULL ht_module, unlike SignalDict subtype. get_module_state_by_def(tp_base) or _PyType_GetModuleState(tp_base) can be used in signaldict_*().