Add LM precompile workload

[sChoiKr]

Summary

Add a TensorKitchen-side precompile workload for the LM frontend so the first cpd(...; solver = :lm) and nncpd(...; solver = :lm) call no longer pays the large operator-path JIT cost at runtime.

What changed

• add PrecompileTools as a dependency
• load precompile.jl from src/TensorKitchen.jl
• add a dedicated src/precompile.jl workload covering:
  • direct LM solve on canonical CP models
  • cpd(...; solver = :lm) with random init
  • cpd(...; solver = :lm, init = :alswarm)
  • nncpd(...; solver = :lm) with random init and ALS warm start
• update CPD/NNCPD pipeline docs to mention :lm in the shared manifold-solver dispatch list
• keep the public default solver on the RGD path

Why

The LM implementation itself was not spending most of its time in TensorKitchen callbacks. The main issue was first-call latency from JIT compilation along the Manopt nonlinear least-squares/operator path. This PR moves that cost into package precompilation so end users do not see a 40-50 second pause on the first LM call.

Validation

Fresh Julia process timings after precompile:

• direct canonical LM solve on a (6,5,4) CP model: about 43s -> 0.15s
• cpd(...; solver = :lm, init = :alswarm) on a (6,5,4) CP example: about 43s -> 0.18s
• nncpd(...; solver = :lm, init = :alswarm) first call: about 0.28s

I did not rerun the full repository test suite in this PR flow.