test(otel): add GPU, Neuron, and EFA DRA-path integration tests

[samehkhalil]

Description of the issue

EFA metrics on EKS can be exposed to pods two ways: the EFA device plugin
(vpc.amazonaws.com/efa) and Dynamic Resource Allocation (DRA), where EFA
devices are allocated via a DRA driver (dranet, driver dra.net) and
ResourceClaims. The agent's OTel Container Insights pipeline supports both, and
the awsdevicepodcorrelation processor has a dedicated DRA code path that watches
ResourceClaims/ResourceSlices via the K8s API and bridges the DRA device
identity (a PCI name, e.g. pci-0000-00-1e-0) to the EFA metric label (e.g.
rdmap0s30) via the dra.net/rdmaDevice ResourceSlice attribute.

Today there is no integration coverage for the DRA path — only the device-plugin
path is exercised. This adds an end-to-end test so the DRA path is validated on a
real cluster: EFA metrics are emitted per device and correctly correlated to the
pods that claim them.

Description of changes

Add an integration test that provisions a cluster exposing EFA via DRA and
validates per-device EFA metric correlation end to end:

• test/otel/multi_efa_dra/ — queries the EFA metrics in CloudWatch and
  asserts the DRA path produces the expected per-device series and pod
  attribution: each device is a distinct series with its EFA attributes
  (aws.efa.device, ENI, port), the device a pod claims is correlated to that
  pod (name/namespace/container), and devices not claimed by any pod carry no pod
  attributes.
• terraform/eks/daemon/otel-multi-efa-dra/ — provisions the cluster,
  installs dranet (eks/aws-dranet) as the DRA driver, deploys the
  observability chart and agent, and runs an efaburn workload that requests one
  EFA through a ResourceClaimTemplate.
• Generator entry wiring the test to its terraform module.
• helm_chart_repo_url variable so the chart can be cloned from a fork when
  validating chart changes not yet merged upstream (defaults to upstream).

Pinned to k8s 1.34 (the rest of the suite is on 1.35): the processor's DRA
informers watch resource.k8s.io/v1beta1, which 1.34 still serves alongside the
GA v1. The generator entry is commented with this rationale; it moves to 1.35
once the processor's DRA client is bumped to v1.

Dependencies (the EKS lane is green only after these land): the chart on
main must render the DRA correlation config (dra_device_types on the
dra.net driver + dra.net/rdmaDevice keying) and grant the agent ServiceAccount
get/list/watch on resource.k8s.io resourceclaims/resourceslices, and the
released agent image must include the DRA processor path. Committed defaults
already point at that merged end-state (public agent image, chart main, upstream
chart URL), so no follow-up edit is needed once those merge.

License

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Tests

Ran the module end to end on a live cluster (EKS 1.34, c6in.32xlarge, EFA via
dranet). All tests passed.

Verified against ground truth:

• The node exposed 2 EFA devices, rdmap0s30 and rdmap0s31.
• efaburn's ResourceClaimTemplate allocated pci-0000-00-1e-0, which maps via
  dra.net/rdmaDevice to rdmap0s30 → correlated to the efaburn pod
  (namespace default, container efaburn).
• rdmap0s31 was unclaimed → no pod attributes.

Each EFA metric (efa_rx_bytes, efa_tx_bytes, efa_rx_dropped,
efa_rdma_read_bytes) reported per device with correct DRA-based pod correlation.
Cluster torn down after the run.