fix(serve): size tensor parallelism from the target instance's GPUs

[AkshayShah03]

Issue #, if available: N/A

Description of changes:

_get_default_tensor_parallel_degree used the GPU count of the machine running the build whenever that machine had any GPUs. It only used the gpu_count passed in when the build machine had none:

tot_gpus = len(available_gpus) if available_gpus else gpu_count

The SageMaker endpoint builds for TGI (NUM_SHARD) and DJL (TENSOR_PARALLEL_DEGREE) pass the GPU count of the target instance, taken from instance_type. So:

• building on an 8-GPU box for a 1-GPU ml.g5.2xlarge set the degree to 8, and the container cannot start;
• building on a 1-GPU notebook for a 4-GPU ml.g5.12xlarge set it to 1.

This change uses gpu_count when it's given, and only probes local GPUs when it isn't, as in local container mode, where the callers don't pass it.

Testing:

• New tests in tests/unit/model_server/test_djl_utils.py: the target count wins in both directions (1 local → 4 target, 8 local → 1 target), and without gpu_count the local GPUs are used. The first fails before this change.
• tests/unit/model_server and test_model_builder_servers_hf_model_id.py: 134 passed.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.