Share ort session between instances

[glebashnik]

This is a rebased and cleaned-up follow-up to #248.

It adds optional sharing of a single ONNX Runtime session across Triton model instances.

This significantly reduces memory growth with multiple instances. GPU throughput can be lower with shared
sessions, so this is intended as an opt-in memory optimization.

The feature is controlled by a model config parameter:

parameters {
  key: "share_session_between_instances"
  value: { string_value: "1" }
}

Once enabled for a model, it applies to all instance groups of that model. However, sharing is within each
instance group, not globally across all instance groups.

The default behavior is unchanged: each model instance creates its own ORT session.

Changes since #248:

• Locking improvement for shared session initialization using RAII.
• Improved instance group name parsing without regex. This handles multi-digit instance group indexes and model
  names with regex special characters.
• Documented the new parameter.
• Added an integration test that verifies session reuse and runs inference.

I also ran local CPU/GPU benchmarks with Triton Server 2.68.0 and a ModernBERT embedder to measure memory use
and throughput. Tested with enable_mem_pattern, do_copy_in_default_stream, and use_ep_level_unified_stream; these did not meaningfully affect performance for shared sessions.