fix(lib-cloudinfo): type CloudProduct.gpusPerVm as Float to keep fractional-GPU counts

[pditommaso]

What this changes

Types CloudProduct.gpusPerVm as Float instead of Integer.

CloudInfo reports fractional GPU counts for fractional-GPU instance families — e.g. AWS g6f.2xlarge reports gpusPerVm=0.25 (fractional NVIDIA L4). Decoding 0.25 into an Integer field truncated it to 0 (Jackson ACCEPT_FLOAT_AS_INT), so a genuine GPU instance became indistinguishable from a CPU-only one for any consumer checking gpusPerVm > 0. The neighbouring fractional fields (memPerVm, onDemandPrice) are already Float; gpusPerVm was the anomaly.

With Float, fractional counts survive and > 0 correctly flags any accelerator.

Evidence (live CloudInfo, amazon/compute/us-east-1)

type	category	gpusPerVm (wire)	Integer (before)	Float (after)
g6f.2xlarge	GPU instance	0.25	0	0.25
g6f.large	GPU instance	0.125	0	0.125
gr6f.4xlarge	GPU instance	0.5	0	0.5
g6.2xlarge	GPU instance	1	1	1.0

Tests

Added CloudProductTest cases asserting a fractional gpusPerVm survives JSON decode (0.25 stays 0.25, > 0), plus whole/absent values. ./gradlew :lib-cloudinfo:test passes.

Downstream note

This is a breaking type change for consumers that auto-unbox getGpusPerVm() to int (notably sched), which need a matching follow-up to consume Float and treat > 0 as GPU. AcceleratorFeatures.hasGpu(features) (token-based) is unaffected.

Closes #97

Impact context

Downstream in sched, the truncation let g6f.2xlarge be mixed into a CPU-only task's instance pool, contributing to a spot stock-out (seqeralabs/sched#789).

🤖 Generated with Claude Code