chore: extract shared intervention-tier scaffolding

[RhizoNymph]

What

Extracts the payload-agnostic scaffolding that the steering and patching tiers each hand-copied (and that feat/clamping would have copied a third time):

• intervention_kernel_common.py — one kernel-warmup harness (normalize_warmup_sizes + run_kernel_warmup with Triton-version-tolerant compiled-variant accounting). The steering, monitor, and patch warmups keep only buffer allocation and a per-size drive(n) closure. Also fixes the patch copy's cache probe, which only read the pre-Triton-3.6 cache attribute and silently reported 0 variants on Triton 3.6.
• intervention_common.py — hook_attrs / derived_attrs for the per-hook buffer attr-name dicts (values are a getattr-by-string runtime contract; now derived in one place), and BufferKnob for config-first buffer sizing with a TEST-ONLY process-global fallback (backs the patch slot knob).
• phase_tiers.py — generic base/prefill/decode PhaseTiers container behind SteeringManager's global vectors; replaces the per-family phase if/elif ladder and clear-triple. The global_*_vectors attribute names survive as properties.
• validate_vector_entries — single vector-spec check core behind both validate_steering_vectors (drain path, reject unknown layers) and _validate_vectors_spec (HTTP path, skip unknown layers under TP). Message texts and strictness are per-caller data, so the two paths can no longer drift on check logic.

Why

Clamping (feat/clamping) and the SAE branches behind it instantiate the same tier template; the copies were already drifting (the Triton 3.6 cache probe above, and the two vector validators disagreeing on unknown-layer handling). This lands the shared scaffolding on feat/integration so those branches adopt it on rebase instead of adding third/fourth copies.

Pure refactor — no behavior change: public names, signatures, tested attribute names, buffer attr-name values, and error texts are preserved.

Note: expect merge conflicts with feat/clamping in steering_manager.py / steering_model_runner_mixin.py and the kernel modules; clamping should rebase onto this and adopt the helpers.