Remove AlignedArray and Aligned Matrix from src and tests

[Anipik]

Fixes #1018

cc @danmosemsft @eerhardt @TomFinley @yaeldekel @shauheen

[TomFinley]

Sounds great to me, thanks @eerhardt .

[Anipik]

@eerhardt @danmosemsft @TomFinley @tannergooding

Method	without align	align	percentage slower
Test_Multiclass_WikiDetox_BigramsAndTrichar_OVAAveragedPerceptron	5.8 s	5.297 s	9.5
CV_Multiclass_WikiDetox_BigramsAndTrichar_OVAAveragedPerceptron	105 s	96.8 s	8
CV_Multiclass_WikiDetox_BigramsAndTrichar_LightGBMMulticlass	310 s	297 s	4
CV_Multiclass_WikiDetox_WordEmbeddings_OVAAveragedPerceptron	310 s	309 s	0
CV_Multiclass_WikiDetox_WordEmbeddings_SDCAMC	192 s	196 s	-2%
TrainKMeansAndLR	242 ms	232 ms	4
MakeIrisPredictions	3 ms	3 ms	0
MakeSentimentPredictions	51 ms	53 ms	-3
MakeBreastCancerPredictions	1.3 ms	1.3 ms	0
Test_Ranking_MSLRWeb10K_RawNumericFeatures_FastTreeRanking	1.4 s	1.1 s	3
TrainTest_Ranking_MSLRWeb10K_RawNumericFeatures_FastTreeRanking	33.91 s	32.7 s	3
TrainTest_Ranking_MSLRWeb10K_RawNumericFeatures_LightGBMRanking	34.33 s	32 s	7

updated the last row

[Anipik]

The test failing is flaky.
I queued another build and it was successful. https://dnceng.visualstudio.com/public/_build/results?buildId=25891&view=results

[eerhardt]

I'm concerned with some of the perf degradation. We should investigate why these tests are that much slower.

[tannergooding]

@eerhardt, I would suspect that many of the tests have small enough inputs that the cost of re-aligning was potentially small/negligible.

The current code also doesn't have the fixes that ensures most of the read/writes are actually aligned (except for the few leading/trailing unaligned elements, if any)

[tannergooding]

The codegen is also going to be worse for the native intrinsic code, since it is emitting "legacy" encoded instructions (and won't be able to fold away the loads anymore).

I imagine the diff will be less for netcoreapp3.0 using the managed HWIntrinsics.

[eerhardt]

The current code also doesn't have the fixes that ensures most of the read/writes are actually aligned (except for the few leading/trailing unaligned elements, if any)

Can we prototype that and see if that fixes the perf regression?

[markusweimer]

What is the status of this PR? I am asking because the last update has been some time ago, and it is now in conflict with master.

[Anipik]

@markusweimer this pr is waiting on #1274 . i have the conflicts resolved in my local branch. probably #1274 will get merged by today

[Anipik]

Closing this PR, will continue the work after my vacation