Optimize getting run plan when using region/instance-type filters

[jvstme]

Current

The user can filter instance offers by region or instance type, e.g.:

dstack run . --region westus3

However, the current implementation always prepares offers for all regions and instance types and applies the filters only after that. This may slow down getting the run plan significantly due to checking quotas and availability for the instances that will then be filtered out.

Proposed

Pass these filters to Compute.get_offers and adjust backend implementations to only prepare offers for the requested regions and instance types.