PLAN: re-prioritize post-M9 milestones (M10–M22)

Replaces the two-item post-M9 plan (conv-pool training + 1.0 release)
with twelve milestones derived from a structured review of the M9
codebase, ordered by user-visible value:

  M10 quantized inference          M16 mixed-precision training
  M11 mlx::fast::* fused kernels   M17 conv-pool training (was M10)
  M12 zero-copy round-trip         M18 observability & telemetry
  M13 EXLA gradient conformance    M19 error surfacing
  M14 serving concurrency cookbook M20 GPU interop pointers
  M15 native linalg                M21 mix emily.doctor
                                   M22 1.0 release (was M11)

Old M10 (window-reductions + small-CNN training) re-slotted as M17
behind the inference, oracle, serving, linalg, and mixed-precision
work — its reach is narrower than "make Bumblebee inference
production-ready".

Also:
- Non-goals: mixed precision moved out (now M16), distributed
  training and native optimizer library remain.
- Core design decision #9: noted unimplemented status; M12 delivers.
- Testing philosophy: Grad row cites EXLA goldens (M13).
- Project decisions: streams + training entries updated to point at
  the milestones that change them.

Author: ausimian