Zero-Copy C Environment Implementation

In high performance reinforcement learning systems, environments written in C/C++ expose their internal state through pre allocated memory buffers. The Python side creates numpy arrays that directly view the environment's internal memory, so each environment step updates these arrays in place without allocating new memory or copying data. This is called zero copy design. However, zero copy semantics introduce a subtle hazard: if a user stores a reference to the observation returned by the environment, that reference is actually a view into the internal buffer. When the environment steps again, the buffer is overwritten and the stored reference silently shows the new data, not the old. Users must explicitly copy data if they want to retain past observations. Implement zero copy env simulation(num envs, obs dim, step data, commands) that simulates a vectorized C environment with pre allocated buffers and processes a sequence of commands. Parameters: num envs: number of parallel environments obs dim: dimension of each observation vector step data: list of tuples (obs 2d list, rewards list, dones list) representing data produced by the C environment at each timestep commands: list of command tuples to execute in order: ("step",) advance the environment by writing the next entry from step data into internal buffers (in place) ("store view", name) store a zero copy reference to the current observation buffer under the given name ("store copy", name) store an independent copy of the current observation buffer under the given name ("read", name) read the current content of the stored snapshot and append it to results ("read buffer",) read the current observation buffer content and append it to results ("check alias", name) check whether the stored snapshot shares memory with the internal buffer; append True/False to alias checks ("auto reset",) for environments marked as done, zero out their observation and reward entries in place (simulating automatic reset in C environments) The internal buffers use float64 for observations and rewards, and int8 for done flags. All buffers are initialized to zero before any commands execute. Return a dictionary with keys: "reads": list of 2D nested lists from read/read buffer operations (rounded to 4 decimal places) "alias checks": list of booleans from check alias operations "total copies": number of store copy operations performed "total views": number of store view operations performed "total bytes saved": total bytes saved by using views instead of copies (each avoided copy would have cost num envs obs dim 8 bytes) "buffer reallocs": always 0 for a correct zero copy implementation