[TF FE] Fix centernet and correct FloorDiv translator for signed integer type

[pavel-esir]

Details:

• Centernet's topk operation returns large int32 values (greater than 1000000), even though they are integer FloorDiv/Div(inp_1, inp_2) + Floor operation is performed in float16 and because of that it causes accuracy problems.

• To solve this need to performs FloorDiv operation in integer with a subgraph:

res = x / y; if x > 0 and y > 0
res = x / y - 1; if (x < 0 xor y < 0) and (x mod y != 0)

• checked on separate bus: no degradations caused.

Tickets:

• CVS-130526

[pavel-esir]

@popovaan @rkazants i reverted removing of abs and instead added a separate tests cases where i statically specify inputs instead of random generation. I found that even for a very simple cases the current main, e.g.
FloorDiv(10, 10) which is equivalent to python 10 // 10 should be 1, but after compression of IR OpenVINO gives answer 0,

because if this denominator 10 is converted to Div(input_1, 10) => Mul(input_1, 1/10) then when 1/10 = 0.1 is compressed to f16 it becomes 0.099975.., which causes difference even for CPU inference!!! (because floor(10 * 0.099..) = floor(0.99..) = 0)

But after this fix works fine, since this constant is stored exactly without compression.

One more thing, since layer test call mo instead of ovc save_model is not called in layer tests, and therefore even with my fix tests appeared red. I added some changes to layer test class so that ovc is called, if these changes are radical i will revert them.

[pavel-esir]

@alvoron FYI since Div for signed ints on Arm returns values different from cpu/gpu this FloorDiv also differ. xfailed Div and FloorDiv layer tests for arm until CVS-132377 is resolved.