Typecast for bfp16 to bfp8 is incorrect

[TT-BrianLiu]

Describe the bug
See #15889 for context.
Previously, our handling of mantissa rounding is always round to nearest. If tied, round up. This is incorrect and in this PR #16997, I will correct it to round to even on ties.

However, after the change, I noticed ttnn.typecast op doesn't give exact match for host conversion vs. device conversion. These 3 tests in tests/ttnn/unit_tests/operations/eltwise/test_eltwise_typecast.py break:

• test_typecast_bf16_to_bfp8_b
• test_typecast_bf16_to_bfp8_b_various_input
• test_typecast_bf16_to_bfp8_b_with_inf_nan

To Reproduce
Steps to reproduce the behavior:

• Will update after PR is merged.
• For now, apply these changes: https://github.com/tenstorrent/tt-metal/pull/16997/files#diff-34586446481f78d399b595901ec11c6120eebcff5359f9500094967701207338L150-R169
• Run the above tests

Expected behavior
The tests should fail without those changes and pass with them if hardware does the same rounding spec (ie. round to even).

[rdjogoTT]

Just confirmed with the HW team that WH and BH both DO NOT support rounding to even, only rounding up in case of ties. This means that the original implementation was accurately modelling HW.