[Core] Add lowest() to platform numeric_limits<float> (issue #3509)

[XFDG]

Summary

• add lowest() to cutlass::platform::numeric_limits<float>
• cover LinearCombinationClamp<float> on both the host and device in the core functional tests

Problem

LinearCombinationClamp obtains its lower saturation bound from cutlass::platform::numeric_limits<ElementOutput>::lowest(). The float specialization exposes max() but not lowest(), so explicitly instantiating LinearCombinationClamp with a float output fails to compile even though float is part of the class's public template contract.

The new member returns -max(), matching the standard definition of the lowest finite IEEE-754 float while reusing the specialization's existing exact max() encoding.

Validation

Using CUDA 13.1 and -arch=sm_100a on an NVIDIA B200:

• current main fails to compile the issue reproducer with numeric_limits<float> has no member lowest
• the fixed source compiles the same host/device probe successfully
• the probe runs and prints host=2 device=2
• git diff --check passes

The core regression test uses the existing unary_operator kernel and HostTensor utilities to exercise both compilation paths.

Fixes #3509