Rewrite pytorch-runners-routing skill for the completed ALI→OSDC migration

**Impact:** dev tooling only (Claude skill docs for PyTorch CI runner routing)
**Risk:** low

## What
Overhauls the `pytorch-runners-routing` skill to reflect the finished
ALI→OSDC migration (pytorch/pytorch #189219 removed the `arc` experiment and
the `use-arc` output) and adds a new `references/old-ali-infra.md` covering the
legacy Lambda-autoscaled EC2 world.

## Why
The skill still described the old four-input model (bare label + `runner_prefix`
+ `use-arc` toggle + `arc.yaml`) and the four-prefix state table, none of which
match current reality. After #189219, ARC-on-EKS is the unconditional
destination for everything the determinator routes: `map_ec2_to_arc.py` runs on
every test-matrix, the determinator only emits dash prefixes (`mt-`/`lf-`/`c-mt-`),
and `_linux-test.yml` no longer has a `test` vs `test-osdc` branch. The rewrite
reframes the whole model around the two infrastructures and the dot-vs-dash
prefix split, so anyone touching `runs-on:` values, `arc.yaml`, or the
determinator gets the current picture.

# Notes
- Doc-only change — no code, config, or workflow is touched, so nothing runs
  differently. Reviewers should mainly sanity-check the factual claims:
  the prefix→infra table, the `meta_only_runners` / identity-passthrough
  behavior (notably `linux.dgx.b200.8` now being a passthrough, not absent),
  and the per-workflow `check_experiments`/`opt_out_experiments` inventory,
  which was re-derived from live source.
- The GPU workflow table dropped a couple of stale rows (`vllm.yml`,
  `operator_microbenchmark_compare.yml` kept; `inductor-pallas` retained) and
  now annotates which builds pin `runner_prefix: "mt-"` — worth a quick glance
  against the actual workflows if you have them handy.

Signed-off-by: Jean Schmidt <contato@jschmidt.me>

Author: jeanschmidt