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

[intelgaoxiong]

Details:

Extends Host Flash Attention (HFA) and Pyramid attention to operate with
the block-split KV cache produced by SplitKVCacheIntoBlocks (Part 1/4).

Section 1 — Shared infrastructure:

• util.hpp/cpp: rename isPastKeyValuesKey/Value → isPastKeyParam/isPastValueParam; add isPastKeyParamContiguous / isPastValueParamContiguous
• sdpa_utils.hpp/cpp: new file extracting shared SDPA parameter utilities (previously duplicated between HFA and Pyramid)
• attention.hpp: extend SDPAIndices with past_key_blocks/past_value_blocks vectors; extend Attention struct with per-variant block indices

Section 2 — Host Flash Attention:

• host_flash_attention.cpp: loop over all Concat inputs in build_sdpa_param_mapping() to collect _past_key_block_indices / _past_value_block_indices
• base_sync_infer_request.cpp: replace scalar past_key/past_value checks with is_past_kv() lambda

Section 3 — Pyramid Attention:

• pyramid_attention.cpp: add is_block_split path in process_pyramid_model() that shrinks each pyramid-variant Concat to idx past blocks; collect_concat_block_indices() helper
• base_sync_infer_request.cpp: add block_mode + bind_block_ports() lambda in bind_pyramid_attention_inputs()
• just_sync_infer_request.cpp: share_kv_block_buffers() for pyramid variants
• partitioning/patterns/sdpa.cpp: relax Concat input-count check for multi-block inputs

This is part 3/4 of the block-based KV cache feature split.

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).