[C API] Continue implementation

[vkuzmin-uber]

Summary:

Continue with basic implementation of a C API for Neuropod. Issue #407

Test Plan:

Added a test that loads a model and fails, runs inference successfully and fails, and verifies the name and platform.

[codecov]

Codecov Report

Merging #409 into master will increase coverage by 0.31%.
The diff coverage is 96.66%.

@@            Coverage Diff             @@
##           master     #409      +/-   ##
==========================================
+ Coverage   89.55%   89.86%   +0.31%     
==========================================
  Files          87       87              
  Lines        5024     5080      +56     
==========================================
+ Hits         4499     4565      +66     
+ Misses        525      515      -10     

Impacted Files	Coverage Δ
source/neuropod/bindings/c/np_tensor.cc	75.00% <50.00%> (-25.00%)	⬇️
source/neuropod/bindings/c/c_api.cc	100.00% <100.00%> (ø)
source/neuropod/backends/tensorflow/tf_backend.cc	90.94% <0.00%> (+2.43%)	⬆️
source/neuropod/internal/neuropod_loader.cc	92.00% <0.00%> (+2.66%)	⬆️
source/neuropod/bindings/c/np_status.cc	81.81% <0.00%> (+27.27%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 20b6e1d...5fdc025. Read the comment docs.

[VivekPanyam]

Looks good overall!

[VivekPanyam]

I'm not sure it's a good idea for status to be optional here (and in the other methods). Doing so might make this interface somewhat error prone.

It's obviously easier to pass in nullptr than to use and check am NP_Status, but I think we should force usage.

When I was working with the TF C API, it was somewhat inconvenient to have to deal with statuses everywhere, but it forced me to do error checking and I think that led to more robust user code (and a better debugging experience).

Any thoughts @selitvin

[vkuzmin-uber]

I thought abut some extra efficiency for Infer. "error handler" instead of status may be more efficient (it would never call it on success), but let's make it required for now.

[VivekPanyam]

One more thing :)

[thuningxu]

Let's make sure we test with a C compiler.

[vkuzmin-uber]

Let's make sure we test with a C compiler.

This is a very good point. Tensorflow also uses C11 to implement smoke test for its C API. Just published update for PR, see test_c_api.c.

[VivekPanyam]

Looks good! Two minor comments and we should be good to go :)

[VivekPanyam]

LGTM!