QNN EP: Add MaxPool and AveragePool tier-based tests

[yuhuchua-qti]

Summary

• Migrate enabled QNN HTP MaxPool legacy coverage to tier-based tests with a 1:1 mapping.
• Add paired session-snapshot and accuracy coverage for 14 HTP AveragePool-family cases: 12 AveragePool and 2 GlobalAveragePool.
• Generate and review 14 AveragePool-family session snapshot goldens locally; generated JSON goldens are intentionally not included in this PR.
• Preserve all legacy integration tests. CPU and GPU legacy cases remain unchanged; this PR neither disables legacy tests nor relocates GPU coverage.

Test Structure

• Pool is layout-sensitive, so the operator coverage uses session snapshots plus paired accuracy tests.
• averagepool_specs.h provides shared specs that distinguish AveragePool and GlobalAveragePool through op_type.
• Shared support-boundary contracts live in pool_op_builder_test.cc: invalid auto_pad, non-unit dilations, rank-2 inputs, and multiple outputs.
• maxpool_test.cc retains the MaxPool-specific HTP rank-5 rejection.
• The shared component checks are supplemental; they are not replacements for legacy integration mappings.

Validation

• python qcom/build_and_test.py lint --target-py-version 3.12
• python qcom/build_and_test.py test --target-py-version 3.12
• Focused PoolOpBuilder and MaxPool component tests: 5/5 passed.
• AveragePool snapshots: 14/14 passed; forced accuracy: 14/14 passed.
• Linux coverage for pool_op_builder.cc: 100% line coverage and 52.64% branch coverage.

Follow-up

• No enabled HTP legacy ceil_mode case is currently mapped for AveragePool; retain it as a session-level follow-up.