fix: Guard FP4 data type before TensorRT 10.8

[suharvest]

What does this PR do?

Guards the two remaining nvinfer1::DataType::kFP4 references in
FusedNvfp4GemmAllReducePlugin::supportsFormatCombination() so the source can
be compiled with TensorRT releases before 10.8, where that enum value does not
exist.

For TensorRT 10.8 and newer, the original FP4/INT8/FP8 acceptance expression is
unchanged. For older TensorRT releases, the existing INT8 and FP8 alternatives
remain available without referencing kFP4.

Scope

• One source file.
• No plugin ABI, serialization, kernel, or model changes.
• No behavior change for TensorRT 10.8+.

Validation

• Focused compile/preprocessor matrix passed for simulated TensorRT 10.3,
  10.7, 10.8, and 11.0 version macros.
• TensorRT 10.7 preprocessing contains no DataType::kFP4 references.
• TensorRT 10.8 preprocessing retains both original FP4 checks.
• An unguarded negative control fails against the pre-10.8 enum as expected.
• git diff --check passes.

The machine used for this focused validation does not have a Jetson
TensorRT/CUDA toolchain, so a complete JetPack 6.2 plugin build is still
requested from CI or device validation.

Closes #141.