feat: add get_gpu_errors tool for ECC and XID error reporting

[pmady]

Summary

Add a get_gpu_errors MCP tool that surfaces GPU error state — ECC memory errors, XID errors, and driver health.

Background

GPUs can have silent errors that degrade inference accuracy or cause crashes. NVML exposes:

• nvmlDeviceGetMemoryErrorCounter — ECC single-bit and double-bit errors
• nvmlDeviceGetTotalEccErrors — total corrected/uncorrected errors
• XID errors via event monitoring

Requirements

• New get_gpu_errors tool returning per-device error counts
• Fields: ecc_single_bit, ecc_double_bit, retired_pages, xid_errors (if available)
• Graceful handling when ECC is not supported (consumer GPUs)

Suggested output

{
  "index": 0,
  "uuid": "GPU-aaaa-1111",
  "ecc_supported": true,
  "ecc_single_bit_errors": 0,
  "ecc_double_bit_errors": 0,
  "retired_pages_single_bit": 0,
  "retired_pages_double_bit": 0
}

Implementation hints

• Add Errors() ([]ErrorInfo, error) to the Collector interface
• Implement in gpu/nvml.go, stub in gpu/stub.go, add to gpu/mock.go
• Register tool in server/server.go following existing pattern
• Add tests in server/server_test.go using mock

Acceptance criteria

• New tool registered and callable
• Graceful fallback for non-ECC GPUs
• Tests using mock (no GPU required)
• README tools table updated

Difficulty: Intermediate — requires understanding NVML error APIs