[Core] Fix numeric_limits constants for low-precision types (issue #3508)

[XFDG]

Summary

Fix three numeric_limits constants whose encoded values do not match the represented low-precision types:

• float_ue8m0_t::lowest(): 0xfe -> 0x00 in both the std and CUTLASS platform specializations
• tfloat32_t::epsilon(): 0x00001000 -> 0x3a800000 (2^-10)
• half_t::epsilon(): 0x1800 -> 0x1400 (2^-10) in both specializations

The existing core tests now pin each corrected encoding.

Rationale

float_ue8m0_t is unsigned, so its lowest finite value is also its minimum encoded value, 0x00 (2^-127), rather than 0xfe (2^127). TF32 and FP16 each carry 10 explicit fraction bits for numeric_limits purposes, so their machine epsilon is 2^-10.

Validation

Built the same standalone CUDA probe against current main before and after this change using CUDA 13.1 and -arch=sm_100a, then ran it on an NVIDIA B200.

Before:

• both UE8M0 lowest() specializations returned 0xfe
• TF32 epsilon() returned 0x00001000
• both FP16 epsilon() specializations returned 0x1800 on the host and device
• 7/7 checks failed

After:

• all 7/7 host/device encoded-value checks pass
• the probe exits with failures=0
• git diff --check passes

Fixes #3508