Cleanup CPU predict function.

[trivialfis]

• Remove predict instance. It's dead code as we can't use it outside of XGBoost even with C++ include.
• Remove unroll. No performance benefit.
• Optimize dense QDM inference.
• optimize data loading by directly copying data to feature vector instead of going through a workspace.

Partially address #10793

The optimization mostly focuses on dense data and the result varies between CPUs:

| Xeon(R) Gold 6128 |            DMatrix |    QuantileDMatrix |
|-------------------+--------------------+--------------------|
| Master            | 27.980122327804565 | 55.665775775909424 |
| PR                |  23.63674759864807 | 30.158272981643677 |

| Ryzen 9 7900X3D |            DMatrix |    QuantileDMatrix |
|-----------------+--------------------+--------------------|
| Master          | 24.764960527420044 | 31.460495710372925 |
| PR              | 22.532921314239502 | 21.412014961242676 |

[trivialfis]

@razdoburdin Could you please help take a look into the optimization?

I'm not an expert in CPU optimization. The changes in the predictor affects the Xeon much more significantly than the Ryzen. If I remove the dense optimization, it adds about 3 seconds to Ryzen, but 20 seconds to the Xeon.

Looking at some profiling results on Ryzen, the bottleneck seems to be in data loading (movss/movl). Would love to get some opinions.

[razdoburdin]

@razdoburdin Could you please help take a look into the optimization?

I'm not an expert in CPU optimization. The changes in the predictor affects the Xeon much more significantly than the Ryzen. If I remove the dense optimization, it adds about 3 seconds to Ryzen, but 20 seconds to the Xeon.

Looking at some profiling results on Ryzen, the bottleneck seems to be in data loading (movss/movl). Would love to get some opinions.

It is hard to give the exact answer without deep investigation of the changes. My hypothesis are:

1. Xeon benefits more from vectorization due to AVX512 support
2. Xeon has much smaller L3 cache, that makes memory access optimizations more critical.

[trivialfis]

@razdoburdin Thank you for sharing, could you please help review the changes in the CPU predictor when you are available?

It is hard to give the exact answer without deep investigation of the changes

Currently, the evaluation might be even more expensive than training for some datasets. Would be great if we could get some help on that.