`Statistics::is_exact` semantics

**Describe the bug**
It is unclear what `Statistics::is_exact` = `false` means. The docs are here:

https://github.com/apache/arrow-datafusion/blob/a578150e63e344fbaa7d13eda58544482dea4729/datafusion/common/src/stats.rs#L34-L37

These state for this case:

> may contain an **inexact estimate** and may not be the actual value

What does "inexact" mean? Some potential definitions (we only consider `Some(...)` fields here!):

- **underestimate:** There are values within the data source that are NOT included within the statistics, i.e. the statistics do NOT cover the whole range. This could happen when you sample statistics from a larger data source.
- **overestimate:** All values from the data stream are covered by the statistics, but the range might be too large. This can happen when some source doesn't fold predicates into the statistics (which in general is pretty hard to do).
- **both:** The statistics are only a rough guide.

I think there is a pretty important difference between "overestimate" and "both", because the former allows you to prune execution branches or entire operations (e.g. sorts in some cases) while the latter can only be used to re-order operations (e.g. joins) or select a concrete operation from a pool (e.g. type of join).

Side note: Due to predicate pushdown it will be pretty unlikely that there will be exact statistics for any realistic data sources.

**Expected behavior**
Clarify behavior.

**Additional context**
Cross-ref #997.

Start with datafusion/common/src/stats.rs at the documented Statistics::is_exact definition, then read the discussion in #997. Determine which meaning of inexact applies to the Some(...) statistics fields and document the resulting guarantees, including how they may be used for pruning or operation selection.

Written by the indexing model from the issue text.