add support for loongarch64

[mogoweb]

Support non-x86 builds, and let the NN tests run on the Eigen backend

Two independent fixes found while building KataGo on loongarch64 with the Eigen backend.
Each is in its own commit. Neither changes behavior on any currently supported platform.

1. Only pass -mfpmath=sse on x86

The GNU/Clang branch of the build setup treated "not ARM" as "is x86":

if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "(arm|aarch32|aarch64)")
  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsigned-char")
else()
  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpmath=sse")

-mfpmath=sse only exists on x86. On loongarch64 (and riscv64, ppc64le, mips64, ...) every
translation unit fails immediately:

c++: error: unrecognized command-line option '-mfpmath=sse'

so the build dies at ~2% with no usable binary. The else() is now an explicit x86 match.

This is safe rather than merely cosmetic: the comment already notes that SSE-based IEEE float
math is the ABI default on x86-64 and that x87 excess precision cannot occur, so the flag is a
no-op there. Nothing loses IEEE compliance by restricting it to x86 — the goal was just to state
the requirement explicitly.

2. Normalize NHWC and FP16 test args for the Eigen backend

The Eigen backend only implements the NHWC float32 path and hard-errors otherwise:

• eigenbackend.cpp:2425 — inputsUseNHWC = false unsupported
• eigenbackend.cpp:2663 (and siblings) — useFP16 = true not supported

But the test scripts pass exactly those. runsearchtestslimited.sh and runsearchtests.sh were
written for OpenCL, so they invoke e.g. runsearchtests <model> false false 0 false (NCHW) and
runsearchtestsv8 <model> true true true (FP16). Every NN test therefore aborted with SIGABRT
before doing any work.

TestCommon::overrideForBackends() exists for precisely this kind of normalization, but it only
knows OpenCL and TensorRT, and testsearchmisc.cpp's runNNSymmetriesTest / runNNBatchingTest
never call it at all. Rather than extend that function (it takes only two params, so it cannot see
useFP16) and fix up ~13 call sites, the clamp goes in the two startNNEval helpers:

• cpp/tests/testsearchcommon.cpp
• cpp/tests/testtrainingwrite.cpp (its own file-local startNNEval)

Those are the single choke point every NN test goes through, so this also covers the two
testsearchmisc.cpp cases that never reached overrideForBackends. Guarded by
#if defined(USE_EIGEN_BACKEND), so OpenCL / CUDA / TensorRT / ROCm / Metal / ONNX builds are
bit-for-bit unaffected.