[XLA] Speed up exhaustive_f32_elementwise_op_test by using our own...

[XLA] Speed up exhaustive_f32_elementwise_op_test by using our own implementation of LiteralTestUtil::Near.

LiteralTestUtil::Near does a bunch of stuff that we don't care about, like
keeping track of mean/stddev.  It also operates on literals not 1D arrays, and
so does expensive index calculations on each entry.

Rather than laboriously try to fix it up, it's easy enough just to reeimplement
the relevant part.

Speedup is ~8/5 = 1.6x.

PiperOrigin-RevId: 231305563