Improve proxy management for data using COW

[pfernique]

Some proxies are used to fasten data manipulation (e.g., WeightedData) are error prone (segfaults) or induce non user-friendly manipulation (e.g., ProxyError thrown).

From now, these proxies are constructed from a pointer to the decorated data and store this pointer.
Since this decorated data can be deleted, some tests are required to ensure that pointed data still exists.
Since most of these proxies are constructed from data that won't be modified, using the COW design pattern will as efficient and far more user-friendly.

[pfernique]

This has to be done for:

• Univariate sample spaces. This has been tackled in pfernique@07b5cb8. Note that only categorical sample spaces are COW since others don't require a lot of copying (2 ints or 2 doubles and 2 bools).
• Multivariate sample spaces. This has been tackled in pfernique@ee0a384 for multivariate sample spaces created from univariate sample space vector.
• Univariate data frames. This has been tackled in pfernique@abd5bba but is not used for any other class using univariate data.
• Multivariate data frames.
• Data maskes.
• Estimations.