Handle hex tensor decoding across compatible value types

[SahilKumar75]

Why

Hex-encoded tensor input currently depends on field cell type in a brittle way. Feeding float-encoded hex to a tensor<bfloat16>(...) field can fail with an out-of-bounds error in JSON tensor parsing.

What

• Added expected-size-aware hex decoding in tensor JSON parsing.
• When hex length matches the expected number of values but a different supported cell width, decode using inferred source type and let normal numeric assignment convert to target field type.
• Applied this in both:
  • vespajlib JSON tensor decoding (JsonFormat)
  • document feed JSON reader (TensorReader)
• Kept empty-string behavior intact with existing error messages.

Tests

• mvn -pl vespajlib -Dtest=JsonFormatTestCase test
• mvn -pl document -Dtest=JsonReaderTestCase test
• mvn -pl document -am -Dtest=JsonReaderTestCase -Dsurefire.failIfNoSpecifiedTests=false test

Regression coverage added

• float-hex -> bfloat16 decoding for dense values
• float-hex -> bfloat16 decoding for mixed block values
• document JSON reader coverage for both direct dense and mixed block hex forms

Fixes #35611

[bratseth]

LGTM!