Add systematic error testing framework (AT-101)

[devin-ai-integration]

Make sure to read the contributing guidelines before submitting a PR

Summary

This PR addresses JIRA ticket AT-101 by expanding test coverage for error handling and edge cases. It creates a systematic error testing framework with three main components:

1. Memory exhaustion testing (test-memory-exhaustion.cpp) - Tests controlled OOM scenarios including small/medium/large allocations, buffer overflow, and recovery mechanisms
2. Invalid input validation (test-invalid-inputs.cpp) - Tests malformed tensor shapes, type mismatches, dimension limits, and parameter validation
3. Backend operations error scenarios - Extends existing test-backend-ops.cpp with error injection capabilities

Key Changes

New Test Files (377 lines total)

• tests/test-memory-exhaustion.cpp - 6 systematic memory pressure scenarios with controlled failure injection
• tests/test-invalid-inputs.cpp - 6 input validation scenarios testing tensor creation edge cases

Error Injection Infrastructure

• ggml/src/ggml-alloc.c - Adds environment variable-based error injection:
  • GGML_ALLOC_FAIL_THRESHOLD - Fail allocations >= specified byte size
  • GGML_ALLOC_FAIL_COUNT - Fail after specified number of allocations
  • Non-intrusive design that only activates when env vars are set

Extended Backend Testing

• tests/test-backend-ops.cpp - Adds GGML_TEST_ERRORS environment variable to enable error injection test cases
• tests/CMakeLists.txt - Registers new tests using existing llama_build_and_test pattern

Testing Results

• All 38 existing tests continue to pass (0 regressions)
• New tests pass all 12 combined scenarios (6 memory + 6 input validation)
• Error injection properly integrates with existing tensor operations

Important Review Areas

🔴 Critical: The error injection mechanism modifies production code (ggml-alloc.c). While gated by environment variables, reviewers should evaluate whether this approach is acceptable or if a different testing strategy should be used.

🟡 Thread Safety: The error injection uses a static counter variable that could have race conditions in multi-threaded scenarios.

🟡 Test Isolation: Tests rely on global environment variables - verify that parallel test execution won't cause interference.

🟢 Integration: CMake integration follows existing patterns and all tests are properly registered.

Human Review Checklist

• Evaluate if error injection in production code (ggml-alloc.c) is acceptable architecture
• Review thread safety of static alloc_count variable in ggml_alloc_should_fail()
• Verify test scenarios actually test meaningful error conditions vs superficial checks
• Confirm environment variable cleanup prevents test interference
• Check if new test files follow existing code style and patterns

Link to Devin run: https://app.devin.ai/sessions/55de95f979c14d71aa4b6b3125a2ccf1
Requested by: Alex Peng (@alexpeng-cognition)

JIRA: AT-101

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring