Add datatype-aware numeric conversion reads

[janosh]

Problem

H5Dataset::read_raw::<T> intentionally exposes the stored representation and accepts a target based on element width. A consumer that wants semantic numeric values in one Rust type must instead inspect DatatypeMessage, dispatch across every supported integer or floating-point storage type, call read_raw with that exact type, and convert each value.

For example, reading arbitrary floating-point datasets as f64 requires separate f32 and f64 branches. Reading arbitrary integer datasets as one wide checked type requires branches for signedness and widths from 8 through 64 bits. Generic metadata and scientific-data readers duplicate this dispatch.

Proposed API

Please consider a separate conversion API that leaves read_raw unchanged, for example:

let floats = dataset.read_numeric_as::<f64>()?;
let integers = dataset.read_numeric_as::<i128>()?;

A corresponding hyperslab form would also be useful:

let values = dataset.read_numeric_slice_as::<f64>(&starts, &counts)?;

Requested semantics

• Inspect the stored datatype class, signedness, byte order, and width rather than only comparing element sizes.
• Support the standard primitive integer and floating-point datatypes.
• Perform checked integer conversions and report overflow instead of wrapping.
• Define float-to-integer and narrowing-float behavior explicitly, or reject those conversions initially.
• Reuse the existing decoded data path so contiguous, chunked, and filtered datasets behave consistently.
• Keep read_raw as the zero-conversion representation-level API.

Even an initial pair of focused helpers such as read_f64 and read_i128 would remove substantial downstream datatype dispatch while keeping conversion policy explicit.