Support fractional GPUs in the AWS Accelerators module

[jnioche]

What

Lets the CCF Accelerators module handle instance types that get a fraction of a physical GPU, and adds a script to check the bundled GPU data against what EC2 actually offers.

Code — every numeric value read from the configuration now goes through Number instead of a cast to Integer, which accepts both the Integer and the Double Jackson produces for 1 and 0.125:

• quantity is a double, so a fractional share scales the wattage through the existing amount * energy_used * quantity / 1000;
• min/max watts are double too. This fixes a latent ClassCastException: GPU_INFO already holds "max": 76.5 for NVIDIA_TESLA_P4 and would have thrown the moment an instance type referenced it;
• gpu_utilisation_percent is a double read as a Number, matching the pattern in OperationalEmissions, so "gpu_utilisation_percent": 12.5 is now valid config.

Data — added the five fractional sizes AWS offers, all NVIDIA L4:

instance type	quantity
g6f.large, g6f.xlarge	0.125
g6f.2xlarge	0.25
g6f.4xlarge, gr6f.4xlarge	0.5

Taken from the AWS G6f spec table and confirmed against the GpuPartitionSize returned by ec2 describe-instance-types. The URL is recorded in the file's sources.

Script — scripts/compare_gpu_accelerators_aws.py pages through aws ec2 describe-instance-types, keeps everything with a GpuInfo, and reports instance types missing from accelerators.json (with whether their GPU model already has wattage data), ones the queried regions no longer offer, GPU count mismatches (float-tolerant), GPU model mismatches, and undefined or unused GPU_INFO entries. --region is repeatable since instance types are region-specific; --strict exits 1 on real gaps. Stdlib only, same shape as the existing compare_gpu_accelerators_boavizta.py.

Notes for the reviewer

Scaling a full GPU's min-max range linearly by the partition size means a g6f.large is modelled as drawing an eighth of an idle L4 (1 W). Hardware-partitioned GPUs don't idle down proportionally — the card's floor is shared across partitions — so fractional sizes are under-estimated at low utilisation. Improving that needs a per-partition floor in GPU_INFO, which is a modelling decision beyond this PR.

Running the new script against us-east-1 surfaces gaps that are not addressed here, since each needs wattage figures sourced first:

• the whole g7 family (NVIDIA RTX PRO 4500), g7e (RTX PRO Server 6000) and p6-b300.48xlarge (B300) are missing, and none of those three models has a GPU_INFO entry;
• g5g.* is mapped to NVIDIA_T4 but AWS reports T4g — inherited from CCF, and whether the power profile differs enough to deserve its own entry is a judgement call;
• NVIDIA_TESLA_P100 and NVIDIA_TESLA_P4 are defined but referenced by no instance type.

Full test suite passes (322 tests). AcceleratorsTest covers the fractional sizes in both CUR and FOCUS form, plus a decimal gpu_utilisation_percent.