Fix rounded_from_bits casting in float8.h

[apivovarov]

Currently XLA bazel build //xla:literal_test fails with error (Clang 18)

In file included from xla/literal.cc:16:
In file included from ./xla/literal.h:41:
In file included from ./xla/array.h:36:
In file included from ./xla/types.h:27:
In file included from external/tsl/tsl/platform/ml_dtypes.h:19:
bazel-out/k8-opt/bin/external/ml_dtypes/_virtual_includes/float8/ml_dtypes/include/float8.h:1343:21: error: non-constant-expression cannot be narrowed from type 'WideBits' (aka 'unsigned short') to 'ToBits' (aka 'unsigned char') in initializer list [-Wc++11-narrowing]
 1343 |       bits = ToBits{rounded_from_bits};
      |                     ^~~~~~~~~~~~~~~~~

Solution:

Narrowing conversions in an initializer list, is not allowed in C++11 and later without an explicit cast.

We need to explicitly cast the value to the target type (ToBits) to indicate that we are aware of the potential narrowing

@jakevdp Can you have a look?

[apivovarov]

Seems that it can be merged @jakevdp @hawkinsp

[apivovarov]

This issue impacts TensorFlow/JAX pull request CI builds, for example: tensorflow/tensorflow@f4b0b3e

Jake, Peter, Is it possible to merge it today? @hawkinsp @jakevdp