evaluatePartition() addresses event numbers with integers. It was
changed to use std::size_t. This protects from negative indices, and is
future proof when large numbers of events are available.

o Add a RooSpan that uses a std::span in order to pass around data.
o Change rudimentary batch data interface from std::vectors to spans.
o Clean up iterators in RooVectorDataStore.
o Revert RooNLLVar to single-value computations to have something ready for testing.
o Increment class version of RooVectorDataStore.

Speedup of about 4x achieved, but only ever reads first data column.

- Move functions for finding index of items from RooArgList to RooAbsCollection.
- Fix override/not-yet-implemented warnings in various classes.
- Improve comments.
- Improve readability by using new iterators and correcting indentation.

- Add a RooArgSet to the batch interface that holds variables found in the dataset.
- Using the new BatchHelpers, PDFs can now search for their proxies in the dataset.
- Implement batched/vectorised computations for Gauss and Poisson.

Small cleanups and comments in RooSpan, RooAbsTestStatistic, RooAbsReal.

RooNLLVar now has two modes to switch between scalar and batch+vectorised computation.
The RooNLLVar implementation is not ready, yet. In particular, it's not vectorised.

RooAddPdf allocates its own memory to compute addition of PDFs. The next step
is to take memory in the data store such that computations can be reused.

- Add class BatchData to hold results of batch computations to RooRealVar
- Make batch data accessible through RooRealProxy::getValBatch(begin, end)
- Change batch computation interface in RooAbsPdf, AddPdf, Gauss, Poisson to
getValBatch(begin, end, normSet)
- Make data in RooVectorDataStore accessible to RooRealVar/RooRealProxy
- [Cleanup] Remove traceEvalHook, as nobody uses it
- Update typedefs in std::span backport
- Add conversion constructor from RooSpan<T> to RooSpan<const T>

Replace the existing Kahan summation by a version with multiple accumulators.
This can be auto-vectorised by most compilers when filled from a container with
contiguous memory.

- BatchData: Add possibility to choose whether batches are initialised
- RooAbsReal: Add interface to clear batches between computations
- Adapt already vectorised PDFs to above change
- Implement returning of weight variable in RooVectorDataStore
- Properly enable RooSpan conversion constructor from T to const T
- Add cache value checking for debug builds
- Start refactoring NLL logic into single value and batch branches.
Allows to easily switch between the two.

- [ROOT-9818] Add a fallback implementation for evaluateBatch to
RooAbsReal such that non-batch classes can be used in the same
computation as batched classes.
- Make getValBatch in RooAbsDataStore pure virtual, and place dummy
implementations in derived classes.
- Add code to check batch against scalar computations when compiled in
debug mode.

- Change indexing of batches to a begin+size interface.
- Remove requirement that batches have to be allocated (and their size
given) before the computation. This allows for easier integration with
different data storage backends, which might possibly yield shorter
batches than asked for. This also allows to reduce memory pressure by
not allocating batch storage for every entry in the dataset. Memory can
be recycled for each batch.
- Fix ROOT::span constructor to both take const and non-const vectors.
- Fix override warning in RooDataSet.

- Vectorised the error checking routine in RooAbsPdf, and make sure that errors are communicated
also in batch mode.
- Instead of clearing batches in between NLL runs, only mark them dirty.
- Stabilise error-checking debug routines against inf/nan.

- AbsArgs need a name to be indentified in datasets and PDF trees.
- This commit adds a check such that empty names cannot be used.