Fix Granite residual contribution hooks

[jlarson4]

Description

• Granite-family HF blocks compute residual + sublayer_out * residual_multiplier inline, so blocks.N.hook_attn_out / hook_mlp_out exposed raw module outputs. The value was off by the multiplier and wrong as intervention points.
• New ScaledResidualBlockBridge makes hook_attn_out / hook_mlp_out real HookPoints firing on the scaled contribution. Reads see raw * scale with a bit-exact forward, a write of v (including in-place mutation) reconstructs v as the contribution, and backward hooks always route through the HookPoint.
• Adopted by the granite, granite_moe, and granite_moe_hybrid adapters, with residual_multiplier added to the config passthrough. Zero scale is rejected at construction.
• On hybrid mamba layers hook_attn_out is pruned at bind time. With experts, hook_mlp_out stays absent, the block_sparse_moe + shared_mlp sum has no single capture point.
• Renamed the block capture-hook lifecycle (_maybe_wire_capture_hooks etc.) since it now wires more than pre-LN capture.
• Tests use local tiny fixtures at residual_multiplier=0.22: residual identities, contribution = raw × scale, zero-ablation, write/in-place-mutation/bit-exactness regressions, backward-hook gradient flow, and HF forward parity across dense + MoE + hybrid.
• Tests verified to fail on the unfixed adapter, allowing us to remove the GraniteMoeHybridForCausalLM strict-xfail in the alias-resolution audit.

Fixes #1648 – Granite portion

Type of change

• Bug fix (non-breaking change which fixes an issue)
• This change requires a documentation update

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility