[NPUW] Add block-based KV cache support for HFA and Pyramid attention

[intelgaoxiong]

Details:

What this PR does

Extends Host Flash Attention (HFA) and Pyramid Attention to operate with the block-split KV cache produced by SplitKVCacheIntoBlocks. After that transformation a single past_key / past_value parameter is replaced by N independent block parameters each feeding one input of a multi-input Concat node. Both attention backends must
detect this new layout and adapt their compilation-time model shaping and inference-time tensor dispatch accordingly.

This is Part 3/4 of the block-based KV cache feature series:

Part	PR	Description
1/4	#35012	SplitKVCacheIntoBlocks graph transformation
2/4	#35013	KVCacheBlockManager — block allocation and lifecycle
3/4	this PR	HFA and Pyramid attention: block-KV compilation & runtime
4/4	#35018	LLM inference pipeline integration

---

Changes by section

Section 1 — Shared infrastructure

• util.hpp/cpp — rename isPastKeyValues{Key,Value} → isPastKeyParam / isPastValueParam; add …Contiguous variants that match only single-parameter (non-block) names; used by both HFA and Pyramid to distinguish contiguous from block-split layouts.

• sdpa_utils.hpp/cpp (new file) — extracts SDPAPatternNodes (holding vector<…> past_key_param_nodes / past_value_param_nodes for 1 or N elements), find_sdpa_pattern_nodes(), and find_mask_parameter(). Previously duplicated between HFA and Pyramid; now shared.

• attention.hpp — extend function::Attention with past_key/value_block_variant_param_indices (ordered by Concat input); extend compiled::SDPAIndices with past_key_blocks / past_value_blocks vectors.

---

Section 2 — Host Flash Attention

Compilation (host_flash_attention.cpp):

• build_sdpa_param_mapping() now iterates all Concat inputs (not only the first) to discover and record every block-parameter index into _past_key_block_indices / _past_value_block_indices.
• These indices are promoted into compiled::SDPAIndices.past_key_blocks / past_value_blocks for use at inference time.
• The tile models (HFA_Tile, HFA_Final_Tile) are generated once and remain layout-agnostic; only the tensor sourcing strategy differs at inference time.

Inference (attn_subgraph.cpp):

• Contiguous KV: single concatenated KV tensor; the tile loop iterates context_size / tile_size offsets, slicing or zero-copy-viewing into the tensor at each kv_offset.
• Block KV: Loop over past_key_blocks (one entry per block tensor); the final tile uses present_key_tensor. All tiles are dispatched through the same process_tile lambda regardless of source.

---

Section 3 — Pyramid Attention

Compilation (pyramid_attention.cpp):

• Block-split detection — is_block_split = (past_key_param_nodes.size() > 1) or name does not match isPastKeyParamContiguous().
• Contiguous KV — for each pyramid step: clone model → set KV param shapes to current_past_length → reshape() → validate_nodes_and_infer_types().
• Block KV — for each pyramid step: clone model → call shrink_concat_inputs() to keep exactly model_idx past block inputs in the Concat (model[0] → 0 blocks, model[1] → 1 block, …, model[k] → k blocks) → patch_broadcast_constants + patch_reshape_constants → validate_nodes_and_infer_types → collect_concat_block_indices to populate past_key/value_block_variant_param_indices. Precompute past_key_block_port_map (global index → variant port) and past_key_block_port_set for O(1) lookup at inference time.

Inference (attn_subgraph.cpp, just_sync_infer_request.cpp):

• Contiguous KV — bind_function_input() calls util::view(tensor, param.dim, 0, past_len) to present each pyramid
  variant with a correctly sized KV slice; mask is rebuilt per-variant in prologue().
• Block KV:
  • Setup: alias_block_slots() pre-wires all global block-slot ports on the main request to block_0's buffer as a placeholder, so earlier generic binding code never touches unallocated slots.
  • Binding: bind_function_input() → try_bind_block() looks up past_key_block_port_map[global_idx] to dispatch each incoming block tensor to the correct variant-local port. Variants that expose no port for a given block index (e.g. model[0] with 0 past blocks) consume the call silently without set_tensor.
  • just_sync_infer_request.cpp: share_kv_block_buffers() shares block KV buffers across pyramid variant sub-requests to avoid redundant allocations.

Other:

• base_sync_infer_request.cpp — replace scalar past_key / past_value checks with an is_past_kv()lambda; addblock_mode+bind_block_ports()lambda inbind_pyramid_attention_inputs()`.
• partitioning/patterns/sdpa.cpp — relax Concat input-count guard to allow multi-block inputs.

Tickets:

• EISW-206740

AI Assistance:

• AI assistance used: no / yes
• If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).