Fix test_Meshgrid

[toslunar]

Probably this fixes #8268. (Reopen the issue later if it does not.)
This PR reduces the failure rate of test_Meshgrid.

The outputs are six arrays of the shape (6, 5, 3, 3, 6, 3), which seems too big to test backward with float16, in the failed example.

[Crissman]

Thanks. Were you able to replicate the original errors, and how many times did you run this to confirm ok?

[toslunar]

Were you able to replicate the original errors, and how many times did you run this to confirm ok?

I've not tried it. IMO, the PR removes some bad tests, while weakening tolerance of a test is often a trade-off action where the frequency of failures may be measured first.

The cause seems independent of backend because I found test failures with devices native:0 and cuda:0.

[hvy]

Repeated this change with --count 1000 and it failed once with input_arrs=4, indexing='ij', float16, cudaon the backward test. Maybe 4 is still too high?

[toslunar]

To test indexing, I was not so brave to drop input_arrs=4.

• indexing='xy' (default): 1, 0, 2, ..., N-1
• indexing='ij': 0, ..., N-1

[emcastillo]

LGTM

[emcastillo]

Jenkins, test this please

[toslunar]

Thanks @hvy for measuring the failure rate. I updated the PR comment to keep the issue open.

[chainer-ci]

Jenkins CI test (for commit 1d11f53, target branch master) succeeded!