fix: preserve K-quantized embedding weights

[zjn20030811]

Summary

Embedding::init_params currently converts K-quantized embedding weights to F32 because support_get_rows predates GGML's K-quant GET_ROWS implementations. This can multiply embedding memory by up to four times and trigger out-of-memory failures on constrained devices.

This change:

• keeps Q2_K, Q3_K, Q4_K, Q5_K, and Q6_K embeddings in their stored quantized type;
• retains the existing backend capability probe and CPU staging fallback for backends that cannot execute GET_ROWS for a type; and
• adds a regression test covering the allowlist and allocation path for every K-quant type.

Fixes #1930

Verification

• whitespace and formatting checks pass for the changed files;
• the new regression executable is registered with the project's test runner and
  is built when SD_BUILD_TESTS=ON.

The complete build is left to the repository CI matrix because the local
environment cannot complete a native build.

[zjn20030811]

Validation update: the new regression covers Q2_K through Q6_K plus all previously supported embedding types and checks that allocation preserves the stored quantized type. The change keeps backend capability probing/CPU staging intact, so unsupported accelerators retain their fallback. git diff --check passes; CI can exercise the CMake test target.

@leejet please review when convenient.

[zjn20030811]

@leejet The GET_ROWS allowlist now retains Q2_K–Q6_K embeddings and preserves the backend fallback. The new type/allocation regression covers all K-quants and existing types.