feat: add reusable GGUF linear backend and FP8 support

[xindongliu594]

Summary

Add a reusable GGUF linear backend for NVIDIA GPUs, together with the BlockFP8
linear operator required by the current InfiniLM Route B integration.

What changed

• Added linear_gguf operator registration across:
  • C ABI
  • InfiniCore C++ API
  • NVIDIA CUDA backend
• Added GGUF block decoding support for:
  • Q8_0
  • Q4_K
  • Q5_K
  • Q6_K
• Added two execution paths:
  • register-resident GEMV for decode/small batches
  • tile dequantization + cuBLAS GEMM for prefill
• Added workspace-size and dtype/shape validation.
• Added reusable BlockFP8 linear operator and Python bindings.
• Removed machine-specific cuDNN paths and allocator debug-file writes.

Validation

Built successfully on the latest upstream branch:

• GPU: 2 × NVIDIA RTX 5090
• CUDA: 13.0
• Architecture: SM120
• libinfiniop.so: built and linked successfully
• _infinicore: built and imported successfully
• GGUF C ABI symbols verified:
  • infiniopCreateLinearGgufDescriptor
  • infiniopGetLinearGgufWorkspaceSize
  • infiniopLinearGguf
  • infiniopDestroyLinearGgufDescriptor

The corresponding InfiniLM integration also builds successfully against this
InfiniCore branch.

Compatibility

The new GGUF path is opt-in through the InfiniLM integration and does not change
the default behavior of existing non-GGUF operators.

This PR does not claim bitwise cross-backend equivalence. Model-level strict
matching remains an ongoing InfiniLM validation task.

Review notes

The main reusable component in this PR is linear_gguf. The BlockFP8 changes
are included because they are required by the current InfiniLM integration.