Warn & replace dataframes with non-unique indexes

[dagardner-nv]

• Add has_unique_index & replace_non_unique_index helper methods to MessageMeta
• Create a bunch of unittests
• Only use the index for slicing if it is unique, otherwise use a boolean mask
• The file_type argument of read_file_to_df now has a default value of Auto
• DeserializeStage checks for non-unique indexes and replaces them if needed.

This comes at a performance cost in that the DeserializeStage needs to acquire the GIL in order to check if the Dataframe has a unique index, impacting users who never run into this issue. We could work around this by providing a no-check argument to the stage, or we could do the check in the constructor of MessageMeta when we already have the GIL and can perform the check quite cheapely.

Fixes #689
Fixes #686
Fixes #687
Fixes #286
Fixes #626
Fixes #393

[mdemoret-nv]

Change the destructor of MutableTableInfo to LOG(ERROR) instead of LOG(FATAL).

[mdemoret-nv]

After many additions to this PR, here is the final list of changes.

Breaking Changes

• Changes to MultiMessage
  • Consistency checks are now performed on creation. This will raise errors for invalid offset and count configurations. Such as:
    • meta cannot be None
    • mess_offset must be in the range [0, meta.count)
    • mess_count must be in the range (0, meta.count - mess_offset]
  • Derived classes must define their own __init__ to enforce keyword-only arguments
    • More info here
  • get_meta and set_meta use .iloc instead of .loc
    • .iloc is faster but less flexible. This may cause some issues with uncommon data types
  • Bounds are enforced on get_slice
    • Before, there was no checking on the start/stop parameters. So you could get a slice that was larger than your current one
    • Now, slicing ensures that start is [0, mess_count) and stop is (start, mess_count]. Errors are thrown otherwise
  • All __init__ parameters must be passed by keyword
    • This allows properly creating the right derived types in get_slice and copy_ranges. Before each derived type needed to implement these individually, leading to multiple conflicting implementations
• Changes to MultiTensorMessage
  • Consistency checks are now performed on creation. This will raise errors for invalid offset and count configurations. Such as:
    • memory cannot be None
    • offset must be in the range [0, memory.count)
    • count must be in the range (0, memory.count - offset]
    • count must be >= mess_count
  • If a seq_ids tensor is supplied on creation, it must have absolute message IDs instead of relative. This is to standardize whether seq_ids is relative to mess_offset or not. This also guarantees the above checks will not fail due to incorrect seq_ids. This check amounts to:
    • The first element, seq_ids[offset] must be equal to mess_offset
    • The last element, seq_ids[offset + count - 1] must be equal to mess_offset + mess_count
  • Bounds are enforced on get_slice
    • Similar to the bounds check on MultiMessage, this checks count and offset

Additional Changes

• C++ and Python message classes have significantly improved tests to ensure consistent behavior
  • This was responsible for fixing several issues listed above
  • Tests are also parameterized on cudf/pandas and C++/Python to increase coverage
  • Tests now check non-unique and non-monotonic index types
• Corrent inheritance has been implemented for all MultiInferenceXXX and MultiResponseXXX messages
• New Tensor.from_cupy() and Tensor.to_cupy() to test converting between Python/C++ data structures
• New MultiMessage.from_message() utility class to streamline creating messages from other messages
  • This allows for things like MultiInferenceMessage.from_message(incoming_message, count=10) where all properties from incoming_message will be used to create a new MultiInferenceMessage. Additional supplied keyword arguments, count in the example, will override the values in the from message, incoming_message
  • This should be used as much as possible to set consistent offset and count values
• All MultiMessage implementations now take default values for mess_offset, mess_count, offset, count to make creation easier

[mdemoret-nv]

/merge