Unpack eagerly within packed ops that reshape their outputs

[annxingyuan]

This PR is a quick fix to tensorflow/tfjs#834

Changes

• Eagerly unpack outputs of matMul and im2Row so the final reshape occurs on the unpacked tensor
• Add a test that catches the issue

BUG

---

For repository owners only:

Please remember to apply all applicable tags to your pull request.
Tags: FEATURE, BREAKING, BUG, PERF, DEV, DOC, SECURITY

For more info see: /~https://github.com/tensorflow/tfjs/blob/master/DEVELOPMENT.md

---

This change is 

[caisq]

Thanks!

Reviewable status: 0 of 1 approvals obtained (waiting on @annxingyuan, @nsthorat, @caisq, and @dsmilkov)

---

src/kernels/backend_webgl.ts, line 1645 at r1 (raw file):

new UnpackProgram(input.shape);

Just curious: what are the performance implications (if any) of this change?

[nsthorat]

Reviewed 2 of 2 files at r1.
Reviewable status: complete! 1 of 1 approvals obtained (waiting on @annxingyuan, @nsthorat, and @dsmilkov)

[annxingyuan]

@caisq We need to figure out a way to unpack lazily so adjacent packed ops don't have to unpack / pack their outputs / inputs unnecessarily. But there are no performance implications for now because only matMul is packed.