Pre-load offers catalog

[r4victor]

#3479

• Add a server background task that periodically pre-loads gpuhunt offers catalog, thus avoiding catalog download on query (when getting offers).
• Speed up AWS quotas collection by switching from list_service_quotas to get_service_quota.
• Updated gpuhunt to make catalog.load() thread-safe.

Improvements

Offline providers:

• amddevcloud — 1.71s => 1.34s (-22% – network variability)
• aws — 41.43s => 6.61s (-84%)
• azure — 12.49s => 5.50s (-56%)
• gcp — 13.51s => 5.20s (-62%)
• lambda — 8.75s => 2.18s (-75%)
• nebius — 10.74s => 3.80s (-65%)
• oci — 24.48s => 18.27s (-25%)
• runpod — 9.36s => 0.09s (-99%)
• verda — 9.49s => 2.33s (-76%)

Online providers:

• crusoe — 2.90s => 3.59s (+24% – network variability)
• cudo — 11.16s => 2.14s (-81%)
• digitalocean — 1.69s => 1.74s (+3% – network variability)
• hotaisle — 2.13s => 1.69s (-21% – network variability)
• vastai — 4.99s => 2.69s (-46%)
• vultr — 12.69s => 3.48s (-73%)

Notice online providers also improved although they don't use catalog. This is because catalog.query() used to trigger catalog.load() on the first call (S3 download + CSV parsing), which blocked all providers including online ones. With the catalog pre-warmed at startup, online providers no longer wait for that download. Some online providers were fast just accidentally because query() happened with the catalog already loaded.