Dynamic VM sizing, resource env injection, and job lifecycle controls for GCP Batch

[ksuderman]

PR 3 — Dynamic VM sizing, resource env injection, and job lifecycle controls for GCP Batch

Part of #465.

Summary

Makes the GCP Batch runner size VMs from per-tool resource requests, gives tools the
full VM resources, and adds operational controls for job naming and cleanup. Before
this PR every job used a fixed machine_type, tools ran single-threaded regardless of
the VM, job names could collide, N2/N2D local-SSD counts were rejected, and jobs were
always deleted on cancel (losing Cloud Logging for debugging).

What's included

1. Dynamic VM sizing from cores/mem.
Adds cores and mem fields to GcpJobParams. When either is set,
parse_gcp_job_params() computes an appropriate N2 machine type via a new
compute_machine_type() helper in gcp_util.py (with convert_cpu_to_milli() and
convert_memory_to_mib()), choosing between highcpu/standard/highmem variants.
These helpers are intentionally Galaxy-free so Galaxy's direct
GoogleCloudBatchJobRunner can import them from Pulsar in a follow-up and share one
implementation.

2. Inject GALAXY_SLOTS / GALAXY_MEMORY_MB and set the container compute resource.
On a Batch VM there is no scheduler environment, so $GALAXY_SLOTS fell through to 1
and tools ran single-threaded. When cores/mem are set, these env vars are injected
into the task environment and task.compute_resource is set so the container gets the
full VM instead of Batch's default 2 vCPU / 2 GB.

3. Validate local-SSD count for N2/N2D machine types.
Replaces a bare assert size_gb % 375 == 0 with _validate_ssd_size(), which rounds up
to the nearest 375 GB and adds one more SSD for N2/N2D when the count is odd (these
types require an even count), logging the adjustment instead of crashing.

4. Collision-free job names with a configurable prefix.
Replaces the K8s-style name generator with {prefix}-{job_id}-{unix_timestamp} and
caches it so submit/poll/delete agree. Renames the internal helper
gcp_galaxy_instance_id → gcp_job_id_prefix, resolving the prefix as
job_id_prefix > galaxy_instance_id > "pulsar" so jobs from different Galaxy
instances sharing a project are distinguishable.

5. Optional job deletion and AMQP kill().
Both client kill() paths now honor a delete_batch_job destination param (default
true); set false to retain jobs and their Cloud Logging for debugging. Adds the
previously missing kill() to GcpMessageCoexecutionJobClient and wraps deletion in
try/except so a failed delete doesn't cascade.

Files changed

• pulsar/client/client.py
• pulsar/client/container_job_config.py
• pulsar/managers/util/gcp_util.py

Testing

Verified VMs are sized from TPV cores/mem, tools run multi-threaded with the
injected GALAXY_SLOTS, N2 jobs provision with corrected SSD counts, job names no
longer collide, and jobs are retained when delete_batch_job: false. flake8 clean.

Notes

• Renames the internal gcp_galaxy_instance_id helper to gcp_job_id_prefix; if PR 1
  has merged first this PR may need a trivial rebase. Recommended to merge last of the
  three.
• Future work (separate PR): update Galaxy's
  galaxy/jobs/runners/util/gcp_batch/helpers.py to re-export the sizing helpers from
  pulsar.managers.util.gcp_util so both runners share one implementation.