Keep GFD product label within the label-value length limit

[rajathagasthya]

Summary

gpu-feature-discovery builds the *.product label value by joining the GPU
model with the MIG marker and profile (<model>-MIG-<profile>). For GPUs with
long product names this can exceed the Kubernetes 63-character label-value
limit, and NFD then drops the label entirely — so nvidia.com/gpu.product
never lands on the node and product-based node selectors break.

Example (RTX PRO 6000 Blackwell + 1g.24gb), 67 characters:

NVIDIA-RTX-PRO-6000-Blackwell-Max-Q-Workstation-Edition-MIG-1g.24gb

Change

Bound the generated product value to 63 characters in getProductName (the
single point all product labels flow through, so it covers full-GPU, MIG, the
-MIG-INVALID path, the -SHARED suffix, and both output backends).

The first part is the GPU model; everything after it (MIG, the profile,
SHARED) is discriminating, so only the model is truncated — the profile,
which distinguishes MIG labels, is always preserved. The truncated model has
any trailing -/./_ removed so the value remains a valid label value
(must end alphanumeric). Values already within the limit are returned
unchanged, so existing labels are byte-identical.

Example for the case above:

full   = "NVIDIA-...-Edition-MIG-1g.24gb"   (67, over limit)
suffix = "MIG-1g.24gb"                       (11)
room for model = 63 - (11 + 1) = 51
model  = "NVIDIA-RTX-PRO-6000-Blackwell-Max-Q-Workstation-Edi"   (first 51 chars)
result = "NVIDIA-RTX-PRO-6000-Blackwell-Max-Q-Workstation-Edi-MIG-1g.24gb"   (63)

A +me profile behaves the same and shows the trailing-separator trim:
...-Workstation-MIG-1g.24gb.me (the model's truncation point landed on a -,
which is trimmed).

Notes

• Values change only for hardware where the label is currently absent (because
  NFD was dropping it), so no label applied today changes. The truncated form
  is deterministic and can be used in node selectors.

Resolves #1876