Add multiprocess support

[daveliddell]

Service is refactored to allow for either single process (as it is already), with multiple services in that process, or one process per service.

Default is multi-process! Please provide feedback if you would rather see single process as the default.

• Factored HTTP out of ClientGenerateBatchProcess, replacing it with a generic response_handler callback
• Factored Service creation out of lifecycle, replacing it with a service manager class. There are two service managers: one for single process and one for multi-process
• Added instance_num to service, so that the service can self-identify which instance it is (for diagnostic messages etc.)
• Shortfin and service setup moved to a new LlmServiceEnvironment class, used by both service managers
• Multi-process service manager uses a request and response queue to send request objects to a subprocess, managed by LlmServiceProcess, which returns the raw bytes of the response to the main process
• Multi-process service manager has some refinements to service selection to try to keep batches of 16 together: new requests don't go into a queue until the previous flight has disembarked. Gains ~0.5-0.7 RPS. More work to be done after the merge
• Single process service selection has some minor tweaks, but the multi-process algorithm isn't there yet because it's more invasive to the service class. Will revisit after merge
• generate_request HTTP handler refactored to delegate service selection to service manager. Provides the HTTP-aware response handler to ClientGenerateBatchProcess
• New command line switch to choose single process over multi-process