Fix shallow slice bugs: dtype and mem leak

[dsmilkov]

When doing shallow slicing:

• the dtype associated with the tex data can get out of sync with the dtype of the sliced tensor.
• there is a mem leak related to the intermediate tensor when doing data reads of sliced tensor.

Also fix the chain API for unstack.

Fixes tensorflow/tfjs#1257
Also partially tensorflow/tfjs#569 (fixes Coco-SSD when using tfjs-core 0.15.x once this is cherry-picked)

---

This change is 

[annxingyuan]

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @annxingyuan and @dsmilkov)

---

src/ops/slice_test.ts, line 511 at r1 (raw file):

describeWithFlags('shallow slice an input that was cast', ALL_ENVS, () => {
  beforeAll(() => {
    tf.ENV.set('WEBGL_CPU_FORWARD', false);

Do we need to set this, since jasmine_util.ts specifies WEBGL_CPU_FORWARD to be false for both webgl test environments?

[dsmilkov]

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @annxingyuan)

---

src/ops/slice_test.ts, line 511 at r1 (raw file):

Previously, annxingyuan (Ann Yuan) wrote…

Do we need to set this, since jasmine_util.ts specifies WEBGL_CPU_FORWARD to be false for both webgl test environments?

It's good to set it since it makes the test hermetic (agnostic of what jasmine_util decides). It's also a signal for the reader that to reproduce this problem with casting, you have to not forward to cpu.