refactor(server): move speculative init to speculative.cpp

[wadealexc]

Overview

This PR unifies draft/mtp parameter initialization, model, and context loading. I wrote this primarily considering the perspective of server_context_impl::load_model, after reviewing the main/text model initialization path and comparing against draft/mtp model/context init paths.

Draft/MTP model/context are now initialized almost identically to the main model - a constructor that uses a pimpl and exposes model/context via accessor methods. server_context_impl has been changed so that model_dft and ctx_dft are raw pointers, like the main model.

I also added a helper for common_params initialization that captures the behavior spread across a few different places in load_model and provides a single method that should handle all initialization cases correctly. (see common_base_params_to_speculative)

Additional information

A few things I want to point out:

1. When loading the draft model, the master branch does not set params_dft.n_outputs_max = params_base.n_parallel. Every other use of params_dft in load_model does (both fit and spec_mtp init), meaning those uses inherit server_n_outputs_max(params_base). This seemed unintended to me (possibly reserving more VRAM than necessary?), so I fixed it in common_base_params_to_speculative.
2. Since draft and mtp initialization now sits in a single branch, I opted to bookend the entire branch with load_progress_callback to capture the prior spec_mtp branch behavior. There might be a cleaner way to accomplish this.
3. When common_speculative_init fails, we now reset the model as well as the context (master only resets the context).

I'm happy to make any adjustments as needed.

Requirements

• I have read and agree with the contributing guidelines: Yes
• AI usage disclosure: Yes:
  • I used Qwen 3.5 27B to help me find/correct all of the locations in server-context.cpp where model_dft and ctx_dft were referenced (they're now raw pointers)
  • I used Claude to help review my fixes