[jseketa]

Fixes #708

The bug:

1. clean_dcache_by_address rounds addr down to a line.
2. It counts the lines from size alone.
3. A range that does not start on a line misses its last line.
4. clean_invalidate_dcache_by_address does the same.

The fix:

1. Added dcache_lines.
2. It rounds the first byte down to its line.
3. It rounds the last byte down to its line.
4. It counts the lines from the first to the last.
5. Both functions use it.

invalidate_dcache_by_address is unchanged.
It requires an aligned addr and size.
For aligned input the old and the new count agree.

Overflow guard:

1. The last byte is addr + size - 1.
2. Its address fits for every range, also one that ends at the top of the address space.
3. Rounding the end, addr + size, up would overflow there.
4. That is a panic in debug builds and a wrong count in release builds.
5. A range that runs past the top stops at the top (saturating_add).
6. It does not wrap around to address 0.

The test, dcache_lines_cover_an_unaligned_range:

1. A 1460-byte range 20 bytes into a line needs 47 lines.
2. Every range of 1 to 200 bytes, at 64 start offsets, is fully covered.
3. None of them gets an extra line.
4. Ranges in the last line of the address space count without overflow.
5. A range past the top stops at the top.

The test fails without the fix:

• With the old count: 46 lines instead of 47.
• With the end rounded up: attempt to add with overflow.

Tested:

• cargo test --all --exclude cortex-m-rt --exclude testsuite --features cortex-m/critical-section-single-core: passes
• cargo clippy --all --features cortex-m/critical-section-single-core -- -D warnings: no warnings
• cargo fmt --all -- --check: clean
• cargo build -p cortex-m --target thumbv7em-none-eabihf: no warnings
• cargo build -p cortex-m --target thumbv6m-none-eabi: no warnings