[QNN:Bugfix] Fix StridedSlice NC4HW4 layout and improve QNNPerf robustness

[lanling-47]

Summary

Fix QNN StridedSlice operator for NC4HW4 tensor format and add null-pointer guards to QNNPerf infrastructure calls.

StridedSlice Improvements

• NC4HW4 support: Remap NCHW ranges to NHWC order that QNN expects when input tensor is in NC4HW4 format
• Optional stride: Handle fromType=1 with 4 inputs (no stride tensor) — default stride to 1
• Crash prevention: Replace MNN_ASSERT with graceful nullptr returns + diagnostic MNN_PRINT logging
• Null-check: Validate StridedSliceParam before accessing

QNNPerf Improvements

• Add null-pointer guards for deviceGetInfrastructure result
• Check infraType and createPowerConfigId before use
• Guard all setPowerConfig / destroyPowerConfigId calls against null function pointers and zero config IDs
• Prevents crashes when HTP perf infrastructure is unavailable (e.g. on some device configurations)

Testing

Tested on Snapdragon 8 Gen 3 and 8 Elite devices with YOLO and MobileCLIP models that use StridedSlice with NC4HW4 tensors.

[wangzhaode]

Thanks for the contribution. The QNNPerf guard changes are useful and the NC4HW4 StridedSlice support is needed, but there are several issues to address before merging:

1. The NC4HW4 path remaps ranges from NCHW to NHWC, but beginMask, endMask, and shrinkAxisMask are still passed in the original NCHW bit order. For example, a channel mask on bit 1 must become the last-axis bit after conversion to NHWC. Please remap every supported axis mask consistently with the ranges.
2. computeRangesType1 removes the previous length assertion without adding runtime validation. If begin/end/axes/stride lengths differ, or an axis is outside [-rank, rank), the current loop can read or write out of bounds. Please validate these inputs and reject unsupported cases cleanly.
3. The ONNX converter can produce fromType=1 StridedSlice with 3, 4, or 5 inputs. This PR supports only 4 or 5. Please support the 3-input form where both axes and steps are omitted, or explicitly explain why it must remain unsupported.
4. The unconditional parameter MNN_PRINT in the Creator will log for every StridedSlice during normal inference. Please remove it or guard it with QNN_VERBOSE.
5. Please initialize mPowerConfigId to 0 in QNNPerf.hpp so all early-return and destruction paths have deterministic state.
6. This branch is 72 commits behind master. Please rebase and preserve the unequal Slice fix from [QNN:Bugfix] Fix some bugs when run mmpose model #4626, which modifies the same file.

Please add focused device validation for NC4HW4 slicing on N/C/H/W axes, begin/end/shrink masks, 3/4/5-input forms, negative axes, malformed parameter lengths, and unavailable QNN performance infrastructure. The current environment here has no connected QNN device, so these findings are based on static review.