fork: add hydra-dev CMake preset + forward-decl for hydra_rpc_bridge

[ddvnguyen]

Summary

Two small changes that unblock fast local fork iteration:

1. CMakePresets.json (+18 LOC): new hydra-dev preset —
   same 86;120 fat-arch pair and correctness flags
   (CUDA, cuBLAS, FA, RPC, NVML, NATIVE) as the deploy build, but
   skips LTO and wires in ccache for C/C++/CUDA. BUILD_SHARED_LIBS=ON
   so the new tools/llama-engine/hydra_rpc/ module can resolve
   symbols at link time. Cuts a single-file rebuild on the
   sm_86+sm_120 fat build from ~120–180s down to a few seconds once
   ccache is warm.

2. tools/server/server-context.cpp (+5 LOC): doc-comment +
   forward-declaration for the extern "C" hydra_rpc_bridge
   trampoline. The new tools/llama-engine/hydra_rpc/ module
   (added in PR fork: Phase 1 — merged Hydra RPC server in tools/llama-engine/hydra_rpc/ #37) takes the address of this function and needs
   the declaration visible without dragging in the heavy
   server-context.h header.

Why this is safe

• The preset addition is purely additive — no existing presets
  touched, no flag defaults changed, no other CMake code affected.
• The forward-declaration is also purely additive (the function
  definition immediately follows). It's a documentation
  improvement, not a behavior change.
• Total fork-side delta: 23 insertions across 2 files, no
  deletions.

Cherry-pick vs net-new

• CMakePresets.json: net-new (no upstream equivalent).
• server-context.cpp: the forward-declaration is new (PR fork: Phase 1 — merged Hydra RPC server in tools/llama-engine/hydra_rpc/ #37
  added the function definition but did not add a forward-decl;
  this is the missing piece for the new module to take the
  function's address without including server-context.h).

Cross-repo

• Parent issue: docs: add hydra-dev CMake preset workflow for fast fork iteration hydra_vortex#400 (the parent PR that
  documents the preset in DevelopmentRunBook.md and
  docs/build-environment.md)
• Design: llama-engine: per-GPU llama-engines + merged Hydra RPC + no-weight-transfer COMBINE #36 (v4 design handoff, Phase 1)
• Parent submodule bump PR opens after this merges (per the
  push-before-PR rule in the parent repo's 02-implement.md)

Test plan

• cmake --preset hydra-dev -DCUDAToolkit_ROOT=/opt/software/cuda/13.2 -DCMAKE_CUDA_COMPILER=/opt/software/cuda/13.2/bin/nvcc configures successfully
• cmake --build build-hydra-dev --target llama-engine -j$(nproc) builds cleanly
• ccache -s shows cache hits incrementing across rebuilds
• The forward-declaration compiles (i.e., the new
  tools/llama-engine/hydra_rpc/hydra_rpc.cpp module can
  take &hydra_rpc_bridge without including
  server-context.h)
• Fork CI green (hydra-fork build job)