Add backend capability registry for test harness

Add backend capability registry for test harness

Summary

Add a backend capability registry to enable capability-driven test selection across ExecuTorch backends. Complements the tolerance registry in #19910.

Problem

• No unified way to query what a backend supports (dtypes, dynamic shapes, quantization)
• Skip logic is spread across individual flow files (skip_patterns, param_skip_reasons) with inconsistent patterns
• Adding a new dtype or capability check means editing each backend's flow separately

Proposal

Add backends/test/harness/capabilities.py:

• BackendCapabilities dataclass (supported_dtypes, supports_dynamic_shapes, supports_quantization, max_tensor_rank)
• BACKEND_CAPABILITIES registry covering 16 backends
• should_skip_test(backend, dtype, ...) API returning a skip reason or None
• Integration into TestFlow.should_skip_test() so tests auto-skip with consistent messages

Scope

This will:

• Add the registry and lookup API
• Integrate with TestFlow.should_skip_test()
• Migrate Vulkan dtype skips and ARM dynamic shapes skip to registry

Out of scope:

• Per-operator capability tracking (stays in backend partitioners)
• CI artifact generation

Depends on #20014 (tolerance registry) landing first since they share harness/__init__.py.

Related Issues

• #19910 — Tracking issue
• #13347 — Umbrella discussion on delegate testing strategy

cc: @GregoryComer

Start by reading backends/test/harness/capabilities.py and the existing TestFlow.should_skip_test() implementation, along with the tolerance registry dependency in #20014. Define the registry and lookup API for the listed capabilities, integrate it into TestFlow, and migrate the Vulkan dtype and ARM dynamic-shape skips while covering the 16 backends described in the issue.