fix: refactor kvcache structure and support kernel_block_size

[SJTUGavinLiu]

Background & Motivation

This PR systematically refactors the KVCache data structure and related attention operator interfaces to address two core issues:

1. Per-layer view inconsistency in hybrid cache scenarios: The original KVCache struct stored all layers' KV caches as a monolithic tensor. In MHA/MLA hybrid layouts, each layer may have a different shape and cannot be directly indexed by layer id, forcing callers to handle this with ad-hoc logic.
2. Coupling between KV management granularity and attention kernel granularity: Attention operators (FlashInfer, TRTLLM, etc.) have preferences on block size, but the KV manager's physical block size could not be adjusted independently, limiting optimization opportunities on the kernel side.

---

Key Changes

1. KVCache Data Structure Refactoring

• Introduced a standalone LayerKVCache struct as a per-layer KV cache view, holding kv_cache_base, kv_scale_base, seq_size_per_block, and layer_id for a single layer.
• Removed the legacy monolithic kv_cache_base / kv_scale_base tensor fields from KVCache, replacing them with per-layer kv_cache_base_by_layer / kv_scale_base_by_layer vectors. Added metadata fields: num_kv_heads, head_dim, use_mla, kv_lora_rank, rope_head_dim.
• getLayerCache(idx) now returns a LayerKVCache, automatically reshaping raw 2D buffers in hybrid cache mode: MHA layers → [block_num, 2, kv_heads, seq_size, head_dim], MLA layers → [block_num, seq_size, lora_rank + rope_dim].

2. Introduce kernel_block_size to Decouple Logical and Kernel Blocks

• Added kernel_seq_size_per_block to CacheConfig, configurable via --kernel_seq_size_per_block launch argument or KERNEL_SEQ_SIZE_PER_BLOCK environment variable. When smaller than seq_size_per_block, each physical KV block is split into multiple kernel blocks at the operator level.
• Refactored BlockIds with a dual-index mechanism: block_indices (physical blocks) and kernel_block_indices_ (kernel blocks) are always kept in sync. A kernelBlocks() accessor is exposed, and all mutation operations (add(), remove(), swap(), setAt(), resize(), popBack()) automatically update both index arrays, replacing direct raw vector manipulation.
• KVCacheResource and BatchKVCacheResource interfaces extended to propagate blocks_per_kv_block and group_types.

[xinfei-shi]

如果在非full的这层，blocks_per_kv_block_ 直接定为 1，是不是就不用每个地方都要进行 分支处理了。

[SJTUGavinLiu]

是可以的，那我去掉 is_full 这个标记

[xinfei-shi]

这个是否可以 kv cache manager的get layout接口直接提供 layer_attn_types 这个字段。

[SJTUGavinLiu]

和 group_types 一样做到 CacheConfig 理好了，CacheLayout 也加上这个字段。

[xinfei-shi]

这里还挺奇怪的，专门打印了layer0的shape，要不就别打印了吧，因为不同层的shape还不一样。

[SJTUGavinLiu]

测试的时候加的，我检查一下都去掉