Return pos 0 instead of aborting when IFields data is truncated

[Zhuoxi2000]

Fixes #1048

What / why

ReadVisitor::operator()(IFields&) in io/fields.cc pushes onto end_, but it returns without popping when pos + num_u32 > span.size(). As a result, ~ReadVisitor fails HWY_ASSERT(end_.empty()) and the process aborts. ReadResult in io/fields.h says this case should return pos == 0 instead ("the span is shorter than the data says it should be"). Because of the abort, the model_store.cc diagnostics for corrupt configs and MatPtr blobs ("Error deserializing config", "Deserializing MatPtr %s failed") are never printed.

This PR pops end_ in that branch, the same way the SkipField() branch just above does. It also adds FieldsTest.TestTruncatedData, which reads a span that is shorter than the stored size and expects pos == 0.

Callers that already abort on pos == 0 behave the same as before, except that they now print their own error message.

Testing

cmake --preset make -DCMAKE_BUILD_TYPE=Release -DGEMMA_ENABLE_TESTS=ON -DCMAKE_POLICY_VERSION_MINIMUM=3.5
cmake --build build -j8 --target fields_test blob_store_test
./build/fields_test

• Before (new test only, io/fields.cc from dev @ 0224dcb): aborts with SIGABRT (exit code 134).

  [ RUN      ] FieldsTest.TestTruncatedData
  Invalid IFields: pos 1 + num_u32 112 > size 2
  Abort at fields.cc:110: Assert end_.empty():
  

• After: [ PASSED ] 10 tests.
• blob_store_test: 2/2 passed. gemma/configs_test.cc is Bazel-only; built separately against the CMake build, it passes 9/9.
• clang-format (repo .clang-format) reports no changes to io/fields.cc or to the new test.

Environment: macOS 26.5, Apple M4, Apple clang 17.0.0, CMake 4.4.3.

[jan-wassenberg]

Thanks :)