End-to-end integration tests (PyTorch → ir2c → compile → run)

## Context

Sub-issue of #58.

**Blocked by:** #62 (ir2c)

## Goal

Full pipeline integration tests proving the system works end-to-end.

## Test Scenarios

- [ ] **Training E2E**: PyTorch model → `torch2ir` → shape inference → `apply_training_provider` → `ir2c` → compile C code → run on host → verify loss decreases
- [ ] **Inference E2E**: PyTorch model → `torch2ir` → shape inference → `apply_inference_provider` → `ir2c` → compile → run → verify output matches PyTorch reference
- [ ] **All quantization types**: Float32, SymInt32, Asym tested end-to-end
- [ ] **Numerical correctness**: generated code output matches PyTorch reference within tolerance

## Acceptance Criteria

- [ ] At least one model (e.g., Linear → ReLU → Linear → Softmax) tested E2E for training
- [ ] At least one model tested E2E for inference-only
- [ ] All three quantization types pass
- [ ] Tests run in CI