Load external data in HQQ and RTN quantization passes

[Lidang-Jiang]

Summary

Fixes #2223

After the onnx-ir migration, ir.load() no longer loads external data by default. This caused OnnxHqqQuantization to produce invalid output models when the input model stores weights as external data — the original weight tensors retained stale external data references pointing to non-existent files in the output directory.

Changes:

• Add ir.external_data.load_to_model(ir_model) after load_ir_model() in both OnnxHqqQuantization and OnnxBlockWiseRtnQuantization, consistent with the pattern already used in extract_adapters.py (line 104)
• Add regression tests for both passes with external data models

Before

```
Created model with external data (12.0 MB)
Input model validation: PASSED

--- olive quantize --algorithm hqq ---
Exit code: 0
Output model validation: FAILED
Data of TensorProto ( tensor name: W1) should be stored in /tmp/tmpkh7vp8nn/quantized_hqq/model.onnx.data, but it is not regular file.

--- olive quantize --algorithm rtn ---
Exit code: 0
Output model validation: PASSED
```

After

```
Created model with external data (12.0 MB)
Input model validation: PASSED

--- olive quantize --algorithm hqq ---
Exit code: 0
Output model validation: PASSED

--- olive quantize --algorithm rtn ---
Exit code: 0
Output model validation: PASSED
```

Unit tests (13 passed)

```
test_hqq_quantization_pass PASSED
test_hqq_quantization_pass_produces_valid_output_when_model_has_external_data PASSED
test_rtn_quantization_pass_matmul[True] PASSED
test_rtn_quantization_pass_matmul[False] PASSED
test_rtn_quantization_pass_gather[True] PASSED
test_rtn_quantization_pass_gather[False] PASSED
test_rtn_quantization_pass_produces_valid_output_when_model_has_external_data PASSED
test_rtn_quantization_with_exclusion PASSED
test_rtn_quantization_gather_8bit[True] PASSED
test_rtn_quantization_gather_8bit[False] PASSED
test_rtn_quantization_gather_quantize_axis_forced_to_last_dim PASSED
test_rtn_quantization_shared_gather_weights PASSED
test_rtn_quantization_removes_unused_initializers PASSED

======================== 13 passed, 2 warnings in 2.78s ========================
```

Test plan

• E2E: olive quantize --algorithm hqq on model with external data produces valid output
• E2E: olive quantize --algorithm rtn on model with external data produces valid output
• Unit test: test_hqq_quantization_pass_produces_valid_output_when_model_has_external_data
• Unit test: test_rtn_quantization_pass_produces_valid_output_when_model_has_external_data
• All 13 existing + new tests pass
• ruff check clean on modified files

[Lidang-Jiang]

@microsoft-github-policy-service agree [company="{your company}"]

[Lidang-Jiang]

@microsoft-github-policy-service agree

[xiaoyu-work]

@justinchuby can you confirm this logic? this PR is for #2223

[justinchuby]

Lgtm if you want all weights in memory