Skip placement groups for AWS Capacity Blocks

[r4victor]

Fixes #4344
Fixes #4346

• On AWS, a cluster fleet provisioned into a Capacity Block no longer gets a placement group, since AWS rejects Capacity Block launches that specify one and already co-locates Capacity Block instances in EC2 UltraClusters.
• Compute.are_placement_groups_compatible_with_reservations(backend_type) is replaced by are_placement_groups_compatible_with_reservation(instance_offer, reservation) so that backends can decide per reservation, and both provisioning paths now call it via run_async because the AWS implementation makes an API call.
• get_reservation() gets an active_only flag so a Capacity Block is recognized before it starts, and no longer fails when AWS omits ReservationType, as it does for regular capacity reservations.
• Runs in a fleet with reservation now launch into that reservation even if the run doesn't set reservation itself, since the default run_job() and the placement group check now use the effective requirements instead of the job spec.