Port over BatchPrefillWithRaggedKVCache to HIP

[rtmadduri]

This PR ports over the BatchPrefillWithRaggedKVCacheKernel to HIP. The PR makes changes to

• Remove magic numbers and use defined variables like HALF_ELEMS_PER_THREAD or NUM_ACCUM_ROWS_PER_THREAD
• Adjusts sizes for s_frag, o_frag m and d
• Modifies the shared memory indexing logic

[Copilot]

Pull Request Overview

This PR ports the BatchPrefillWithRaggedKVCache kernel from CUDA to HIP, improving code maintainability by replacing magic numbers with defined constants and adapting shared memory indexing for AMD GPUs.

Key changes:

• Introduces BatchPrefillHandler class and wrapper functions for HIP
• Adds comprehensive test coverage for batch prefill operations on HIP
• Replaces hardcoded values with named constants like HALF_ELEMS_PER_THREAD and NUM_ACCUM_ROWS_PER_THREAD
• Implements platform-specific shared memory indexing for HIP vs CUDA

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 9 comments.

File	Description
libflashinfer/utils/flashinfer_prefill_ops_hip.h	Adds BatchPrefillHandler class and wrapper functions for ragged and paged KV cache operations on HIP
libflashinfer/tests/hip/test_batch_prefill.cpp	New comprehensive test suite for batch prefill operations with various configurations
libflashinfer/include/flashinfer/attention/generic/prefill.cuh	Updates kernel implementation with named constants, platform-specific indexing, and pragma once directive

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.