feat(eks): add GB300 EFA layout and GPU-conditional worker taint

[njhensley]

Fixes #

Proposed Changes

• GB300 EFA layout — derive gpu_family=gb300 from the p6e-gb300 / p6e-gb300r instance-type prefixes and emit a card-0-only interface layout (primary interface + one efa-only, both on network card 0). p6e-gb300r.36xlarge reports network cards 1–8 with MaximumNetworkInterfaces=0, so a real ASG launch rejects any interface on those cards — a single EFA on card 0 is the only valid layout. Note that run-instances --dry-run does not enforce the per-card ENI limit and returns false positives; the real ASG launch fails with ENI limits exceeded on Network Card 1. Verified end-to-end on a live p6e-gb300r.36xlarge cluster.
• GPU-conditional worker taint — apply the default dedicated=worker-workload taint only to GPU pools (gpu_family != "na"), so non-GPU (CPU) workers[] pools come out untainted, matching the GKE and AKS providers, which taint only their GPU pools. EKS previously tainted every worker pool unconditionally, forcing CPU workloads to carry a dedicated=worker-workload toleration. An empty taint string renders as --register-with-taints= in user-data (no taints).
• Tests + schema — add terraform test coverage for the GB300 layout/prefix derivation and the GPU/CPU taint behavior; update the accelerator schema description.

Note on the taint change: any existing non-GPU workers[] pool loses its dedicated=worker-workload taint on the next apply (launch-template change → instance refresh). GPU pools and the gb200 layout are byte-identical (no diff).

Release Note

EKS: add GB300 (p6e-gb300/p6e-gb300r) EFA support with a card-0-only single-EFA layout, and apply the default worker-workload taint only to GPU node groups so non-GPU worker pools are untainted (matching GKE/AKS).