fix(flex-dispatcher): gate DeepEP/HybridEP on compute capability, not device name

[XiaohanZhangCMU]

Stack position — independent. Not part of the Kimi-K3 stack (#33 → #34 → #27) and not required by it: K3's golden trainer config uses the alltoall dispatcher, so it never takes this path. Reviewable and mergeable on its own, in any order.

Split out of #27, which is a Kimi-K3 vendoring PR. This is unrelated to K3 and is our own change, not upstream's, so it does not belong in a PR whose contract is "no original code".

The bug

DeepEP / HybridEP eligibility was gated partly on device_properties.name:

device_properties.major in [8, 9] or device_properties.name.startswith(("NVIDIA B200", "NVIDIA B300"))

name is a marketing string and does not reliably identify the SKU. Our B300 hosts enumerate as NVIDIA L20D, so that startswith test is false on every one of them. DeepEP was skipped and the model silently fell back to the alltoall dispatcher - a throughput regression with nothing in the logs pointing at the cause, because the warning it emits reads as if the GPU genuinely were unsupported.

The fix

Gate on compute capability, which comes from the driver and cannot be relabelled:

• 8 = Ampere (A100, sm_80/86)
• 9 = Hopper (H100/H200, sm_90)
• 10 = Blackwell (B200 sm_100, B300 sm_103)

The tuple was repeated at four call sites across apply_flex_dispatcher_backend and validate_flex_dispatcher_backend (two of them written as not x in [...]). It is now one named constant, _FLEX_DISPATCHER_CC_MAJORS, documented where it is defined - so the next reader does not have to ask what 10 means.

Scope

Behaviour changes in exactly one direction: Blackwell GPUs that previously failed the name check now pass. No compute capability that used to be accepted is now rejected - 8 and 9 were already accepted by the major test, and any B200/B300 matching the old name check is CC 10.

Not exercised on hardware here; this is a gating predicate, and the four sites are covered by the surrounding warning/raise paths. Kimi-K3 itself does not take this path - its golden trainer config leaves moe_dispatcher unset - which is precisely why it was wrong to carry this in the K3 PR.