[peterschmidt85]

Offers can report VRAM that differs from the GPU SKU's known capacity. This makes expected-VRAM filters unreliable and splits equivalent offers into separate groups when grouping by GPU.

• Normalize full-GPU, per-device VRAM against KNOWN_ACCELERATORS, accepting a correction only when exactly one capacity matches the existing ±7% window.
• Keep reported VRAM for unknown models or unmatched/ambiguous capacities, with warnings identifying the offer and expected capacities. Partial allocations keep their assigned VRAM.
• Correct DigitalOcean's combined-VRAM handling and the MI325X registry capacity, and add common RTX models found in real offers.

Examples, in GiB per GPU:

GPU	Backend	Before	After
B300	AWS	268.59	270
B300	OCI	262.5	270
B300	RunPod	288	270
B300	Verda	268	270
L40S	AWS	44.70	48
MI325X, 8 GPUs	DigitalOcean	2,048	256

Compatibility:

• SSH GPU discovery uses the existing helper; shim formats and configuration stay unchanged.
• Kubernetes continues using the registry for GPU memory; MI325X detection now reports 256GB instead of 288GB. No dstack code or configuration changes are required.
• Slurm syntax and P100/A100 mappings are unchanged. In gpu_partitions, gpu: MI325X:288GB will fail to resolve; change it to gpu: MI325X:256GB. Bare gpu: MI325X needs no edit and now resolves to 256GB.