BUG: Make mixed integer and floating-point comparisons exact - #32157

PR summary

Closes #32071.

Comparisons between NumPy integers and real floating-point values currently
use common-dtype promotion. When the floating-point dtype cannot represent
every value of the integer dtype, that promotion can round the integer before
the comparison. For example, np.int64(2**53 + 1) can compare equal to
np.float64(2**53), even though the two scalars have different hashes.

This PR:

• adds exact mixed comparison loops for signed and unsigned 64-bit integers
  against double and long-double values, in both operand orders and for all six
  comparison operations;
• routes legacy ufunc resolution, modern DType promotion, and NumPy scalar
  rich comparisons through those loops when the integer precision exceeds the
  floating-point mantissa;
• preserves scalar-subclass comparison overrides;
• covers representability boundaries, fractional values, signed zero, NaN,
  infinities, both operand orders, scalar and array calls, and the
  equality/hash invariant; and
• makes existing tests explicit where their expected values previously relied
  on lossy mixed integer/floating-point comparison.

After this change:

>>> a = np.int64(2**53 + 1)
>>> b = np.float64(2**53)
>>> a == b
np.False_
>>> a > b
np.True_

Validation performed locally:

• spin build --clean -j 4
• spin test --no-build -- -q: 47,470 passed
• spin test -m full --no-build -- -q: 50,812 passed; the only remaining
  case was denied POSIX shared-memory access by the sandbox and passed when
  rerun outside the sandbox
• spin lint
• git diff --check
• direct reproduction of BUG: numpy int64/float64 comparisons are inexact #32071 and randomized comparisons against Python
  semantics

First time committer introduction

Hi, this is my first contribution to NumPy. I worked on #32071 to make mixed
integer/floating-point comparisons follow Python's exact comparison semantics
and preserve the equality/hash contract.

AI Disclosure

OpenAI Codex was used to investigate the comparison dispatch paths, generate
the C, C++, and Python implementation changes, add and adapt tests, and run
local validation. I reviewed all generated and AI-assisted changes, understand
the implementation and tests, and take responsibility for this contribution.