Thank you. What I needed was exporting an array of floats on one system in the textual form, like C++ initializer list for example and then importing it in another system (Arduino Mega 2560) from C++ initializer list, without losing some information. It looks that byte-copying float to int32_t and vice versa will work just fine.
By the way - I have also noticed that regardless of starting the calculation from exactly the same array of floats (binary the same) after thousands of floating point operations the result on Arduino Mega is not exactly the same as on larger computer. I guess this may be caused because hardware internally always calculates with 8 bytes double.